FORM 2
THE PATENTS ACT, 1970
(39 of 1970)
&
THE PATENT RULES, 2003
COMPLETE SPECIFICATION
(See Section 10 and Rule 13)
Title of invention:
METHOD AND SYSTEM FOR MULTI TARGET TRACKING (MTT) IN A DISTRIBUTED SENSOR (DS) NETWORK
Applicant
Tata Consultancy Services Limited A company Incorporated in India under the Companies Act, 1956
Having address:
Nirmal Building, 9th floor,
Nariman point, Mumbai 400021,
Maharashtra, India
Preamble to the description
The following specification particularly describes the invention and the manner
in which it is to be performed.

TECHNICAL FIELD
[001] The disclosure herein generally relates to Multi Target Tracking (MTT) and, more particularly, to method and system for MTT in a Distributed Sensor (DS) network.
BACKGROUND
[002] Importance of Multi-Target Tracking (MTT) in a Distributed Sensor (DS) network has tremendously grown over the years. Applications like automated warehouse, autonomous driving vehicle, traffic monitoring and area surveillance and the like require tracking multiple targets and their trajectories over the time. The targets in such application are indistinguishable and having random characteristics along with environment noise (clutter) which often generates false measurement data.
[003] Conventionally, MTT applications were developed with centralized fusion technique, where gathering and processing of measurements from all sensor nodes of the DS network is performed centrally and then the resulting estimates are distributed back to the sensors. The conventional approaches have large communication overhead and high risk of single point failure. Development of decentralized computing and information fusion (DCIF) architectures with characteristics like scalability, flexibility, fault tolerance and robustness has attracted the attention of research community. Another challenge in MTT is scenarios where the MTT systems utilize DS to cover larger area of interest and the targets to be identified are mostly undistinguishable. Such scenarios arise typically in flood rescue operation, where multiple sensors (multiple drones with cameras) are used to search and locate multiple targets such as stranded humans and animals in a flood affected area. Thus, for MTT in the DS network knowledge about the targets or identification features is generally not known beforehand, resulting in higher chances of false target detection. Further, due to the restricted sensing region or Field of Vision (FoV) of each sensor (each

drone), the sensors need to collaborate such that area of interest is covered effectively. The major challenges faced by the MTT in DS network are 1) finding tracks from a noisy measurement 2) associating tracks with the targets 3) association of targets sensed by different sensors 4) fusion of inter-sensor associated targets information where there cross correlation between the sensors is unknown. In search and tracking applications, target prior information is mostly unknown.
[004] A state-of-art Gaussian Mixture based Probability Hypothesis Density (GMPHD) filter is a single scan (single sensor) MTT approach to track and estimate the time-varying number of targets. However, usage of the GMPHD filter has limitations and challenges in tracking unknown targets in DS network where inter-sensor information fusion is required. The GMPHD filter approach for MTT in DS faces challenges in terms of 1) inter-sensor data association 2) estimating cross-covariance matrices for information fusion.
[005] Other single sensor MTT approaches in cluttered environment are Bayesian filters developed either using single scan technique to track fixed number of target like joint probabilistic data association filter (JPDAF) or efficient multi-scan methods like multiple hypothesis tracking (MHT), Global Nearest Neighbor (GNN) and Markov Chain Monte Carlo method. But, these existing approaches are computation expensive in many practical implementations. Recent developments provide a MS-MTT method using Particle filter based probability hypothesis density (PHD) developed. However, sharing particle information across sensor (nodes) is communication and computation expensive. Further, existing approaches are often developed with assumption that the number of targets is known and intra-sensor target data association is known. In existing MS-MTT approaches, a further assumption is that all FoVs of sensor are overlapping sensing region and do not provide insights on handling non-overlapping sensing regions. However, in practical scenarios, such as in example above, FoVs of few drones may overlap while that of other drones may not overlap. Further, the existing MS-MTT approaches have avoided considering spawning target and require prior knowledge of birth target

intensities.
[006] Thus, in practical scenarios, accuracy of MTT with existing approaches will have considerable deviation from actual number of targets in the total coverage area, covered by the multiple sensors.
SUMMARY
[007] Embodiments of the present disclosure present technological improvements as solutions to one or more of the above-mentioned technical problems recognized by the inventors in conventional systems.
[008] For example, in one aspect, there is provided a processor
implemented method for Multi Target Tracking (MTT) in a Distributed Sensor
(DS) network. The method comprising, at each sensor among a plurality of
sensors in the DS network, capturing, by a sensor data acquisition model
implemented by the processor, a set of targets lying within a Field of Vision
(FoV) of the sensor from multiple targets present in a total area covered by the
plurality of sensors. The captured set of targets comprise surviving targets
represented by previous target state estimates (Xit-1 ) computed for a preceding
time instance and a plurality of birth targets spawned at a current time instant.
Further, the method comprises dynamically determining, by a birth target
estimator implemented by the processor, the plurality of birth targets from the
captured set of targets based on an adaptive birth Gaussian component estimation
approach. Each birth target among the plurality of birth targets is associated with
a mean (mγ,t ) representing a vector having information of current target state
estimates and features observed at the current time instant, a preset birth weight
(ωy,t) and a state co-variance. Furthermore, the method comprises obtaining, by
a target estimator implemented by the processor, local target state estimates for
the sensor providing a set of local targets by applying a Gaussian mixture based
probability hypothesis density (GMPHD) filter. The GMPHD filter utilizes the
mean (mγ,t ), the state covariance of each birth target and the previous target state

estimates (Xit-1), wherein the local target state estimates provides local
probability densities of local targets estimated by the sensor. Further, the method comprises exchanging, by the target estimator implemented by the processor, with a set of neighboring sensors among the plurality of sensors, the local target state estimates and neighboring target state estimates of each neighboring sensor among a set of neighboring sensors. Furthermore, the method comprises processing, by the target estimator implemented by the processor, the local target state estimates and the neighboring target state estimates over multiple iterations based on a Gaussian Mixture General Covariance Intersection (GM-GCI) pair wise fusion approach to arrive at a final target state estimates, wherein the final target estimates provides an overall probability density function of final targets in the DS network.
[009] In another aspect, there is provided a system, alternatively referred
as a Distributed Sensor (DS) network, for Multi Target Tracking (MTT). The DS
network comprising a plurality of sensors, wherein each sensor in the DS
network comprises a memory storing instructions; one or more Input/Output
(I/O) interfaces; and a processor coupled to the memory via the one or more I/O
interface, wherein the processor is configured by the instructions to capture a set
of targets lying within a Field of Vision (FoV) of the sensor from multiple targets
present in a total area covered by the plurality of sensors, wherein the captured
set of targets comprise surviving targets represented by previous target state
estimates (Xit-1 ) computed for a preceding time instance and a plurality of birth
targets spawned at a current time instant. Further, the processor is configured to dynamically determine the plurality of birth targets from the captured set of targets based on an adaptive birth Gaussian component estimation approach, wherein each birth target among the plurality of birth targets is associated with a mean (mγ,t) representing a vector having information of current target state
estimates and features observed at the current time instant, a preset birth weight (ωγ,t) and a state co-variance. Further, the processor is configured to
obtain local target state estimates for the sensor providing a set of local targets by

applying a Gaussian mixture based probability hypothesis density (GMPHD) filter, wherein the GMPHD filter utilizes the mean (mγ,t), the state covariance of
each birth target and the previous target state estimates (Xit-1 ), wherein the local
target state estimates provides local probability densities of local targets estimated by the sensor. Furthermore, the processor is configured to exchange with a set of neighboring sensors among the plurality of sensors, the local target state estimates and neighboring target state estimates of each neighboring sensor among a set of neighboring sensors. Furthermore, the processor is configured to process the local target state estimates and the neighboring target state estimates over multiple iterations based on a Gaussian Mixture General Covariance Intersection (GM-GCI) pair wise fusion approach to arrive at a final target state estimates, wherein the final target estimates provides an overall probability density function of final targets in the DS network.
[0010] In yet another aspect, there are provided one or more non-transitory machine readable information storage mediums comprising one or more instructions which when executed by one or more hardware processors causes a method for Multi Target Tracking (MTT) in a Distributed Sensor (DS) network. The method comprising, at each sensor among a plurality of sensors in the DS network, capturing, by a sensor data acquisition model implemented by the processor, a set of targets lying within a Field of Vision (FoV) of the sensor from multiple targets present in a total area covered by the plurality of sensors. The captured set of targets comprise surviving targets represented by previous target state estimates (Xit-1) computed for a preceding time instance and a
plurality of birth targets spawned at a current time instant. Further, the method comprises dynamically determining, by a birth target estimator implemented by the processor, the plurality of birth targets from the captured set of targets based on an adaptive birth Gaussian component estimation approach. Each birth target among the plurality of birth targets is associated with a mean (mγ,t) representing
a vector having information of current target state estimates and features observed at the current time instant, a preset birth weight (ωγ,t) and a state co-

variance. Furthermore, the method comprises obtaining, by a target estimator implemented by the processor, local target state estimates for the sensor providing a set of local targets by applying a Gaussian mixture based probability hypothesis density (GMPHD) filter. The GMPHD filter utilizes the mean (mγ,t),
the state covariance of each birth target and the previous target state
estimates (Xit-1 ), wherein the local target state estimates provides local
probability densities of local targets estimated by the sensor. Further, the method comprises exchanging, by the target estimator implemented by the processor, with a set of neighboring sensors among the plurality of sensors, the local target state estimates and neighboring target state estimates of each neighboring sensor among a set of neighboring sensors. Furthermore, the method comprises processing, by the target estimator implemented by the processor, the local target state estimates and the neighboring target state estimates over multiple iterations based on a Gaussian Mixture General Covariance Intersection (GM-GCI) pair wise fusion approach to arrive at a final target state estimates, wherein the final target estimates provides an overall probability density function of final targets in the DS network.
[0011] It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention, as claimed.
BRIEF DESCRIPTION OF THE DRAWINGS
[0012] The accompanying drawings, which are incorporated in and constitute a component of this disclosure, illustrate exemplary embodiments and, together with the description, serve to explain the disclosed principles:
[013] FIG. 1 illustrates Multi Target Tracking (MTT) in an example Distributed Sensor (DS) environment comprising a DS network for tracking a plurality of targets, in accordance with an embodiment of the present disclosure.
[014] FIG. 2 illustrates an exemplary block diagram of a sensor in the DS network, in accordance with an embodiment of the present disclosure.

[015] FIG. 3A and FIG. 3B illustrate an exemplary flow diagram of a method for the MTT in the DS network using the plurality of sensors of the DS network of FIG. 1, in accordance with an embodiment of the present disclosure.
[016] FIG. 4 A through FIG. 8 B illustrates graphical representations depicting results of simulation setup, in accordance with an embodiment of the present disclosure.
DETAILED DESCRIPTION OF EMBODIMENTS
[017] Exemplary embodiments are described with reference to the accompanying drawings. In the figures, the left-most digit(s) of a reference number identifies the figure in which the reference number first appears. Wherever convenient, the same reference numbers are used throughout the drawings to refer to the same or like parts. While examples and features of disclosed principles are described herein, modifications, adaptations, and other implementations are possible without departing from the spirit and scope of the disclosed embodiments. It is intended that the following detailed description be considered as exemplary only, with the true scope and spirit being indicated by the following claims.
[018] Embodiments of the present disclosure provide a method and system, alternatively referred as DS network, for Multi Target Tracking (MTT).
[019] Referring now to the drawings, and more particularly to FIGS. 1 through 8 B, where similar reference characters denote corresponding features consistently throughout the figures, there are shown preferred embodiments and these embodiments are described in the context of the following exemplary system and/or method.
[020] FIG. 1 illustrates Multi Target Tracking (MTT) in an example Distributed Sensor (DS) environment comprising a DS network 100 (alternatively referred as system) for tracking a plurality of targets, in accordance with an embodiment of the present disclosure.

[021] The method disclosed extends a Gaussian mixture based probability hypothesis density filter (GMPHD), which is a closed form MTT solution to overcome limitations of state of art GMPHD filter. The existing GMPHD approach assumes targets birth intensity to be known a prior. But in most practical MTT applications, such as searching and locating unknown targets, it is not possible have knowledge of targets apriori. The method disclosed addresses this problem using an adaptive birth Gaussian component estimation for dynamically determining a plurality of birth targets, which are spawn at a current time instant. The adaptive birth Gaussian component estimation is performed at each sensor among a plurality of sensors of the DS network. Further, based on the estimated birth targets for the current time instant and the existing targets, a number of targets in Field of vision (FoV) of each sensor are determined. The adaptive birth Gaussian component estimation enables dynamically detecting possible birth elements in the FoV of the sensor and confirming or validating the detected birth elements as birth targets. This feature enables dynamically estimating the birth targets without need to have knowledge apriori, as well as validation reduces false detections due to noise in the captured data. An examples provided below explains possibility of presence of such birth targets or abruptly occurring targets and further why possibilities of false detection may exist.
[022] Consider a coverage area where intruding drones are to be detected. Drones are symmetric, rigid and have humming sound when in operation, but have small dimension. Radar based solution, can consider environment noise, birds or large planes flying far away as drone (target), which increases possibility of false detections. In camera based solution, using thermal or visual sensing, challenges such as illumination change, weather conditions and small object detection noise arise, which further increase possibility of false detection. Further, a general assumption during target estimation is that an intruder can breach the protected area (coverage area) only from the boundary and birth intensity indicating possibility of new target detection, is uniformly distributed along the perimeter of the coverage area. Thus most of the exiting

techniques that are developed based on uniform birth intensity concept. However, for accurate estimation of the targets it is necessary to consider other situations where the intruder (target) is already in the coverage area, with perimeter uniform birth intensity such target’s birth intensity may be very low. Further, with a uniform birth intensity over the coverage area can give birth to false detection as targets. In such scenarios the birth estimation methods such as the state of art GMPHD filter that assumes the birth intensities to be known aprioiri may not provide accurate target estimation. Thus, the adaptive birth Gaussian component estimation disclosed id enable to dynamically determining the birth targets, which are spawn at any location within the coverage area.
[023] Once target estimation for each of the sensors is performed, multi sensor information fusion is performed among all the sensors of the DS network to aggregate the individual target estimates and target states and arrive at a total number of targets in a total coverage area of the DS network. Unlike existing multi-sensor information fusion techniques that require knowledge of a cross covariance matrix of the sensors, the method disclosed utilizes a Gaussian Mixture General Covariance Intersection (GM-GCI) pair wise fusion approach that enables fusion without the need to have knowledge of cross covariance matrix of the sensors. As depicted in FIG. 1, the sensors 102-a, 102b and 102-c of the DS network 100 with respective FoVs, alternatively refereed in the FIG. 1 as coverage area a, coverage area b and coverage area-c, effectively cover a total coverage area (a+b+c) for the DS network 100. The sensors, for examples can be cameras mounted on a drone or can be cameras mounted on any predefined positions such as in ware houses, surveillance areas and the like. In an embodiment, the sensors in the DS 100 can be heterogeneous with different types of sensors such as Radars, array of mics, cameras (thermal/ images), 3-d lidars, heat sensors and the like. Applications of the method and system 100, disclosed herein include traffic monitoring system, improving autonomous driving system by using near-by inter-vehicle communication, localizing animals in forest and the like.
[024] As depicted in FIG. 1, of the example of system 100, the coverage

area-a, and coverage area-c overlap and both the sensor 102-a, and the sensor-102c capture target 1 and target 3 and target 4 respectively along with a common target 2. This indicates, target 2 is considered by both the sensor 102-a, and the sensor 102-c due to overlapping FoVs. Further, the sensor 102-c captures target 5 and has no overlap of FoV with any other sensor of the DS network 100. However, the GM-GCI pair wise fusion utilized herein to combine individual target estimates of each sensor to arrive at total targets enables accurate target estimation, irrespective of overlap or non-overlap or size of overlap region, which is not static and may vary. The DS network 100 disclosed herein that utilizes the adaptive birth Gaussian component estimation has applicability in real life or practical scenarios and has enhanced usability providing high accuracy for MTT in any DS network. The reason being, in most real life scenarios or practical applications of MTT, wherein say DS networks are used for tracking of stranded humans or animals in a flood affected area, the targets to be searched and located are not known apriori. Further, the sensors such as drones are not positioned statically and are on move in search of targets. The sensors FoVs is not static, thus the overlapping/non overlapping region varies along with the neighboring sensors. Some FoVs may have overlap while others may not overlap. Further, the overlap and non-overlap status of FoV may change in time.
[025] The functional modules residing in each sensor of DS network 100 are explained in conjunction with FIG. 2 and functionalities of the modules are explained in conjunction with a flow diagram of FIG.3A and 3B respectively.
[026] FIG. 2 illustrates an exemplary block diagram of a sensor 102-a in the DS network 100, in accordance with an embodiment of the present disclosure. It is to be understood that each sensor such as sensor 102-a, 102-b, 102-c and the like have similar modules and functions performed and individual end and in collaboration with neighboring sensors of the DS network to arrive at consensus over total number of targets and target state estimates in the DS network 100. However, for simplicity and brevity explanation is provided with respect to sensor 102-a as an example.
[027] In an embodiment, the sensor 102-a includes one or more

processors 204, communication interface device(s) or input/output (I/O) interface(s) 206, and one or more data storage devices or memory 202 operatively coupled to the one or more processors 204. The one or more processors 204 may be one or more software processing modules and/or hardware processors. In an embodiment, the hardware processors can be implemented as one or more microprocessors, microcomputers, microcontrollers, digital signal processors, central processing units, state machines, logic circuitries, and/or any devices that manipulate signals based on operational instructions. Among other capabilities, the processor(s) 204 is configured to fetch and execute computer-readable instructions stored in the memory. In an embodiment, the sensor 102-a can be implemented in a variety of computing systems, such imaging devices (cameras) with processing capability, laptop computers, notebooks, hand-held devices, workstations, mainframe computers, servers, a network cloud and the like having target sensing capability.
[028] The I/O interface device(s) 206 can include a variety of software and hardware interfaces, for example, a web interface, a graphical user interface, and the like and can facilitate multiple communications within a wide variety of networks N/W and protocol types, including wired networks, for example, LAN, cable, etc., and wireless networks, such as WLAN, cellular, or satellite. In an embodiment, the I/O interface device(s) can include one or more ports for connecting with other sensors such as sensor 102-b and sensor 102-c in the DS network to one another or to another server. The I/O interface 206, through the ports is configured to receive input in terms information target parameters captured by the sensors 102-a, wherein the target parameters are recorded as state dynamic variables such as position, distance, angle of targets and the like. In an embodiment, the target parameters can be image features such as color, shape and the like.
[029] The memory 202 may include any computer-readable medium known in the art including, for example, volatile memory, such as static random access memory (SRAM) and dynamic random access memory (DRAM), and/or non-volatile memory, such as read only memory (ROM), erasable programmable

ROM, flash memories, hard disks, optical disks, and magnetic tapes. In an embodiment a plurality of modules 208 can be stored in the memory 202, wherein the modules 208 may comprise a sensor data acquisition model 212, birth estimator 214 and GMPHD 216, when executed by the processors (s) 2104 are configured to perform MTT based on the adaptive birth Gaussian component estimation and the GM-GCI pair wise fusion. The functions of the modules 208 are explained in conjunction with a method 300 of FIG. 3A and 3B. The memory 202 may further comprise information pertaining to input(s)/output(s) of each step performed by the modules 208 and methods of the present disclosure. The memory may also comprise a database 210 to store the individual target state estimates and number of targets and well as the total or final number of targets and final target state estimates.
[030] FIG. 3a and FIG. 3b illustrate an exemplary flow diagram of a method for the MTT in the DS network 100 using the plurality of sensors of the DS network of FIG. 1, in accordance with an embodiment of the present disclosure.
[031] In an embodiment, the system 100 comprises one or more data storage devices or the memory 202 operatively coupled to the one or more processors 204 and is configured to store instructions for execution of steps of the method 300 by the one or more processors (alternatively referred as processor(s)) 204 in conjunction with various modules of the modules 208. The steps of the method 300 of the present disclosure will now be explained with reference to the components or blocks of the sensor 102-a, as depicted in FIG. 2, other sensors 102-b and 102-c of the DS network 100, as depicted in FIG. 1 and the steps of flow diagram as depicted in FIG. 3A and 3B. Although process steps, method steps, techniques or the like may be described in a sequential order, such processes, methods and techniques may be configured to work in alternate orders. In other words, any sequence or order of steps that may be described does not necessarily indicate a requirement that the steps be performed in that order. The steps of processes described herein may be performed in any order practical. Further, some steps may be performed simultaneously.

[032] Assumption is that the MTT DS environment consist of n targets (final number of targets) and m sensors (plurality of sensors). Each of the sensors, for example the sensor 102-a can measure target parameters such as the state dynamic variables, for example, can be position, distance, angle, velocity, acceleration and the like) of targets or in an embodiment the target parameters can be image features such as color, shape and the like. Each of the sensors, for example the sensor 102-a, is configured to process acquired local data and exchange data with the neighboring sensors (for example, sensor 102-b and sensor 102-c). Each target state evolution and output model is governed by a non¬linear system given as:
x+=ft(x;w);y=ht(x;v ) (1)
Where x+ is next state, x is current state, v and w are the random process
disturbance vectors. The target and measurement state vectors are represented as
an unordered finite set of random state vectors, referred as Random Finite Set
(RFS). Mathematical representation of the DS network 100 is an undirected
graph (G= ( Q, ε ) where Q is the set of sensors nodes (alternatively referred as
plurality of sensors) and ε ⊆ Q × Q is the set representing their connection links.
In particular, (i,j) ∈ ε means sensor node i can receive or send data from node
and is considered as neighboring sensors. For a fixed sensor DS network, the
connection (i,j ) remains intact throughout the operation while for nonstationary
or dynamic sensor DS network the connection ( i, j )is function of inter-sensor
distance.
[033] At step 302 of the method 300, the sensor data acquisition model
212 implemented by the processor (s) 204 of each sensor such as the sensor 102-
a, is configured to capture a set of targets lying within a Field of Vision (FoV) of
the sensor 102-a from multiple targets (say n targets as mentioned above) present
in the total area covered by the plurality of sensors. The captured set of targets
comprise surviving targets represented by previous target state estimates (Xit-1 )
computed for a preceding time instance and a plurality of birth targets spawned at
the current time instant. The variable
symbol X is aternatively represented aqqs X .

[034] At step 304 of the method 300, the birth target estimator 214
implemented by the processor 204 is configured to dynamically determine the
plurality of birth targets from the captured set of targets based on the adaptive
birth Gaussian component estimation approach. Each determined birth target
among the plurality of birth targets is associated with a mean ( mγ,t ) representing
a vector having information of current target state estimates and features
observed at the current time instant, a preset birth weight (ωγ,t ) and a state
covariance. The observed features may be target characteristics such visual
characteristics or sound based characteristics. The state covariance refers to a
birth covariance (Pγ,t ) which depends on a state noise of the birth target when the
target parameters are recorded in terms of the state dynamic variables. In an embodiment, when the target parameters are recorded in terms of the image features, the state covariance provides spread of bounding box and scale changes. As an example, the description herein refers to target parameters in terms of state dynamic variables and is not to be construed a limitation. Thus, the mathematical equations are provided accordingly, considering the target parameters to be state dynamic variables.
[035] Dynamically determining the plurality of birth targets from the
captured set of targets based on the adaptive birth Gaussian component
estimation approach comprises a plurality of sub steps. A first sub-step of step
304 comprises recording, by the sensor data acquisition model 212 implemented
by the processor 204, the target parameters corresponding to each target from the
set of targets. For explanation herein the target parameters are considered to be
the state dynamic variables, however can also be recorded as the image features.
A second sub-step of step 304 comprises representing, by the sensor data
acquisition model 212 implemented by the processor, the set of targets and the
recorded target parameters corresponding to each target as a current
measurement set (Zit ) associated with the current time instant. A third sub-step of
the step 304 comprises determining a set of probable birth elements (Zibc,t ) by
discarding measurements from the current measurement set (Zit) , which have

measurement value below a first preset criteria when compared with the previous
target state estimates (Xit-1 ). The first preset criteria is a distance threshold ( Tbc
) when the target parameters are recorded in terms of the state dynamic variables
and a minimum l2-norm of feature when the target parameters are recorded in
terms of the image features. The mathematical explanation for determining a set of probable birth elements (Zibc,t ) is provided below.
[036] Initialized are seven parameters comprising: 1) distance threshold ( Tbc ) to remove previous target state estimates (Xit-1),
alternatively referred as surviving targets (Xit-1 ) having correspondence in the
current measurement set (Zit )
2) an birth measurement size (δ)
3) a detection uncertainty size (∆)
4) a birth target surety minimum set size threshold ( Tnb )
5) the preset birth weight ( ω γ,t )
6) the state covariance in terms of the birth covariance (Pγ,t )
7) a line model cluster threshold ( T ⊥ ).
[037] At time t, the set of probable birth elements ( Zibc,t ), also referred
as the birth-clutter set (Zibc,t ) having measurements of potential birth targets is
identified and a clutter is formed. Its elements are identified using minimum
Euclidean distance threshold between a measurement (Zit ) and previous target
state estimates (Xit-1 ), It is given by:
Zibc,t=Zit- {zit |min(||Zit-Xit-1||) ˂Tbc } (2)
[038] Referring back to step 304, a fourth sub-step of step 304 comprises generating a birth batch measurement set (Zgr,t ) for the set of
probable birth elements (Zibc,t ) from backward frame sequences of size {detection uncertainity sixze (Δ)+ birth measurement size (δ)}. Thus, to
estimate each birth component or birth element, birth measurement set is defined as:
(Zgr,t)= {Zibc,t-δ- Δ:t ||Zibc,t-δ- Δ:t | = = δ)} (3)

[039] Considered are time steps to account for empty measurement set due to detection uncertainty. If the number of measurement set in Zgr,t is less
than , then no birth target or new target has appeared. Thus referring to a fifth sub-step of step 304, the fifth sub-step comprises selecting or considering birth elements from the birth batch measurement set (Zgr,t) associated with each of
the birth element among the set of birth elements, wherein the selected birth elements cardinality is greater than the birth measurement size (δ).
[040] A sixth sub-step of step 304 comprises determining the plurality of birth targets from the selected birth elements by iterating a plurality of steps for each of the selected birth element. The steps comprise identifying a line model by connecting consecutive measurements in the birth batch measurement set ( Zgr,t)
in accordance with a second preset criteria. The second preset criteria is a maximum possible displacement (Tdis) value set when the target parameters are
recorded in terms of the state dynamic variables and the l2- norm of feature
displacement thresholding when the target parameters are recorded in terms of the image features. Further, identifying the line model comprises adding the selected birth element to a birth cluster (Zbt) if a perpendicular distance (d) of
the measurement corresponding to the selected birth element from the line model is within the line model cluster threshold (T ⊥) and satisfies the second preset
criteria. Furthermore, identifying the line model comprises identifying the added
birth element as a birth target, among the plurality of birth targets to be
determined, if the birth cluster (Zbt ) is greater than the birth target surety
minimum set size threshold (Tnb ). The birth target so identified is associated with
the mean ( mγ,t ) and the features observed at the current time instant, the preset
birth weight (ωγ,t ) and the state covariance, which refers to a birth
covariance (Pγ,t ). The entire steps of the sixth sub-step are mathematically
described below.
[041] Identifying the line model: To estimate the target track
using Zibc,t-δ-Δ:t , considered is birth cluster (Zbt ) with measurement element
zi∈Zbc,t-δ and zj∈Zbc,t-δ+1 used to compute the straight line model. The initial

track elements and combinations can be reduced with constraints of
maximum possible displacement ( Tdis ) between two consecutive measurement.
[042] Forming cluster with line model: The measurement belonging
to:
Zω= Zgr,t/Zbc,t-δ, Zbc,t-δ+1 will be augumented to only if their
perpendicular distance d on the line model is within the line model cluster
threshold (T ⊥ ).
[043] Update birth components and detection for multiple birth targets:
If, |Zbt| ≥ Tnb a new target track is confirmed. An element in Zibc,t which is
member of birth cluster Zbt , is considered as new birth target with the mean
(mγ,t ) representing a vector having information of current target state estimates and features observed at the current time instant, a preset birth weight (ωγ,t) and
the state covariance. The state covariance refers to the birth covariance (Pγ,t ).
Update Zgr,t = Zgr,t/Zbt , which means remove all birth elements of the birth
cluster Zbt from Zgr,t . However, |Zbt| ≥ Tnb means no target track is identified.
These steps are repeated to identify next new target or the birth target.
Pseudo code 1: Adaptive Gaussian birth component estimation
1. Initialize parameters Tnb , δ , Δ, Tbc, ωγ,t, Pγ,t, T ⊥
2. At time t, compute Zbc,t and Zgr,t
3. if | Zgr,t | = = then
4. for zi∈| Zt-δ | do
5. for zj ∈| Zt-δ+1 | do
6. if ||zi-zj || > Tdis then
7. Zbt={zi, zj}
8. compute line model
9. for l i = 1to |Z ω| do
10. Compute ⊥ distance on line
11. Zbt={z∈Zωl1 |d(z) 0.5) are extracted and represent the most likely estimates.
5) Create New Targets: New targets appearing in the environment are accounted either by newborn ones or spawning (single target divide as two different targets).
[051] Referring back to steps of the method 300, upon computation of the local target state estimates by each sensor, at step 308 the target estimator 216 implemented by the processor 204 is configured to exchange with a set of neighboring sensors ( sensor 102-a exchanging with sensor 102-b and sensor 102-c if they are neighbors, which means they are linked with communication link), the local target state estimates and neighboring target state estimates of each neighboring sensor among a set of neighboring sensors.
[052] At step 310 the target estimator 216 implemented by the processor is configured to process the local target state estimates and the neighboring target state estimates over multiple iterations based on the Gaussian Mixture General Covariance Intersection (GM-GCI) pair wise fusion approach to arrive at a final target state estimates. The final target estimates provides an overall probability density function of final targets in the DS network 100. The steps of arriving at the final target estimates comprises fusing a local target with a neighbor target if the local target estimates is present in the received neighboring local target estimates of a neighboring sensor based on the GM-GCI pair wise fusion approach to obtain a set of fused targets. Fusion parameters W are used, which provide intensities associated with the local probability densities of each sensor. Further, the unfused local targets and unfused neighbor targets are marked. Furthermore, an agreement of targets state estimates with neighboring sensor is reached based on the set of fused targets and the marked unfused targets. Post the agreement, the GM-GCI pair wise fusion approach is iterated for all the plurality of sensors at multiple levels of iteration. Thereafter the iterations are terminated

at a level among the multiple levels when an agreement criteria is satisfied to provide the final target state estimates as the overall probability density function.
[053] The mathematical analysis of the GM-GCI fusion approach is provided. Fusing two distributions with unknown cross correlation are performed using covariance intersection (CI). Mahler generalizes CI with exponential mixture for multi-target distribution as:
vW(x)=uaW(x)ub1-w(x)/ ʃvaW(y)ub1-W (y)dy (13)
Where, W is fusion parameter such that 0 ≤ W≤ 1 , va and vb is exponential
mixture model of the two different sensors (a, b). Eq. (13) implies multiply two Gaussian mixture models va and vb which results in having total number of fused
intensities as |va| and |vb| i.e. number of components in mixture would have
increased exponentially and would not even be a Gaussian mixture. Thus approximate method is proposed to compute CI fusion of two sensor estimators using the following elementary properties of Gaussian components.
[054] Product of Gaussian components is a Gaussian component given as:
α1N(x;x2,P1).α2N(x;x2, P2)= α12 N(x;x12,P12) (14)
where, P12 = (P1-1 + P2-1)-1 (15)
x12= P12 (P1-1x1 + P2-1x2) (16)
α12= α1α2N (x1-x2; P1 +P2) (17)
[055] Power of a Gaussian component is a Gaussian component as below.
(αN(x;x,P))w= αWG(W,P)N(x;xP/W) (18)
G(W,P)= [(det(2πPW-1))1/2] / (det (2πP))W/2 (19)

[056] Thus power of the Gaussian mixture model can be written as below
UiW=∑j=1NSjwjWG(W,Pj)N(x;mj,Pj/W) (20)
[057] With above Gaussian properties and eq. (13) the approximated

fused GMM is:

Where, W=w1i / w1i+ wj2 (23)
Pijab = (W(Pia)-1+(1-W)(Pib)-1)-1 (24)
xijab=Pijab(W(Pai)-1 x1i+ (1-W)(Pbj)-1x2j)-1 (25)
αijab=(αia)w(αjb)1-WG(W, Pia)G(1-W,Pjb)N(xia-xjb,0,Pia/W+Pjb/1-W)
(26)
[058] Eq. (21) indicates that exponential increase in the number of
Gaussian components, which is reduced by approximating CI fusion solution for
only those components whose Mahalanobis distance cimdist is below a given
threshold. The Mahalanobis distance cimdist is given by equation (27):
cimdist (xia,xjb=√(xia-xjb)T(Pia/1-W+Pjb/W)-1 (xia-xjb) (27)
[059] Pairwise CI fusion to find overall location probability density function or overall probability density function (Wab, Xab, Pab) from multiple
estimators (sensors) with ( Wia, Xia, Pia),(Wib, Xib, Pib)is computed using Eq. (24)
(26). Here the fusion parameters are the intensities associated with the
probability densities. Fusion of two Gaussian components that are close to each
other is performed while the rest of Gaussian components are kept intact. To
achieve agreement in targets state estimates with neighboring node, iterate the
pairwise fusion algorithm. Algorithm 2 or Example code 2 describes pairwise
fusion of two sensors. {a;b} , where cardinality of sensor-a RFS and sensor-b
RFS is given by and respectively. In steps (5)-(14) of the algorithm 2, all sensor-a targets are compared with sensor-b targets based on
halanobis distance cimdist . Those state estimates whose cimdist is lesser
than threshold Uf are fused. The fused estimates index are augmented in set Ia

and Jb . Compute the fused intensity weights αi,ja,b , mean xi,ja,b and covariance
Pi,ja,b as illustrated in Eq. (23) - (26) and augmented to the respective set. Step (16)-(20) indicates that those target components which are not near to be fused are kept intact.
Pseudo code 2: GM-GCI fusion
1. Procedure ({Wia, xia, Pia}Nas i=1 {Wib, Xib, Pib} Nsb j=1 Uf)
2. αa = [wia] i ∈ [1,Nas]
3. Ia = Jb = x12 = [.]
4. N12 = 0
5. for i = 1 to Nsa do
6. for j=1 to Nsb do
7. if cimdist (xia, xjb) ≤ Uf then ( see eq. 27)
8. Augment indices Ia ∈ Ia Ui and JF ∈ JF Uj
9. Compute W, Pijab, xijab αijab (see eq. 23 to 26)
10. α12 = α12 = U αijab , P12 = P12 UPijab .m12 = m12 Uxijab
11. N12 = N12 + 1
12. end if
13. end for
14. end for
15. W12n= α12n/∑k α12k for n 1 to N12
16. for { i ∈ If } U { j ∈Jf } do
17. m12 = m12 Umia U mjb
18. P12 = P12 U Pia U Pjb
19. W12 = W12 U Wia U Wjb
20. end for
21. Update Wa = W12, ma = m12, Pa = P12
22. end procedure
[060] An architecture of the DS network involves the plurality of
sensors, wherein each sensor, such as sensor 102-a, uses GM-PHD filter with
birth target estimator 214 and reaches in agreement with other sensors over
overall target estimates using the GM-GCI fusion approach as shown in a Pseudo
code 2. At time t, sensor node having a certain sensing region (FoV) searches
for targets. They identify multi-targets that are indistinguishable and stored as
measurement RFS Zt . Zt is used by the birth target estimator 214 to identify a
new target (birth target) with the mean as (my,t ). With previous instant target
estimates (Xit-1 ) and birth-spawn targets, the sensor 102-a, predicts targets state
estimates with a local GMPHD prediction. Zt is also given to GMPHD update
function, which uses Kalman filter with the targets new predicted state to
estimate all the possible target tracks and their intensities are updated. Targets
with weaker intensity as compared to threshold are pruned and remaining target
RFS based on Mahalanobis distance is merged to obtain local target state
estimate. Based on network communication links NH∈ε , they exchange the local
target state estimates with neighbor sensors. At sensor node , the local state
estimated of sensor node belonging to the same target, determined by
Mahalanobis distance cimdist , are fused and merged. Rest of targets which
cannot be fused are kept intact as ith sensor local target state estimates. Pairwise
GM-GCI fusion is done between the neighboring sensors at multiple levels of iteration to reach an agreement on the target state estimation, wherein multiple levels herein is a finite number of iteration (say L times). The final state extraction is done estimate the target states and numbers. If no neighboring sensors are present, i.e. ∑ NH(i;:) = 0 or no communication link between the
sensor 102- and other sensors is present, then the single sensor tracks multi-drones with the proposed birth component estimation algorithm. Thus, if the sensors are not communicating among themselves, then the system works as single sensor MTT. Example pseudo code in the algorithm 3 below provides the end to end steps for MTT in the DS network 100 in accordance with the architecture described.
Pseudo code 3: MTT DS GMPHD pseudo code

1. procedure MS-MTT GMPHD(Node , Time t, Com_link NH)
2. Get Measurement
3. Adaptive Birth Gaussian Component Estimation
4. Local GMPHD Prediction
5. Local GMPHD Updation
6. Pruning and Merging
7. for i = 1, ..., L do
8. for j= 1, ..., Ns do
9. if NH ( i;j ) = =1 then
10. Information Exchange
11. GM-GCI Fusion
12. Merging
13. end if
14. end for
15. end for
16. Estimate Extraction
17. Create Birth and Spawn Targets.
18. end procedure
[061] FIG. 4 through 8 illustrates graphical representations depicting results of simulation setup, in accordance with an embodiment of the present disclosure.
[062] SIMULATION RESULTS: Considered a sparsely connected
sensor network consisting of sensors having a restricted sensing region and can
have an overlapping sensing regions with its neighbors. The setup consist of four
fixed sensors (S1; S2; S3; S4) located at (-4000; 0), (0; 4000), (2000; 2000) and
(4000;-2000) with different sensing region as described in FIG. 4 (a). Here,
considered is rectangular type of sensing region for simulation purpose and it is
valid for other types of sensing regions also. Each sensor can sense targets
position (x,y ) and velocity ( x;y ). A network communication link used here is
given by (28) below:

NH= [ 0 0 0 1 ] (28)
[063] The 2-D environment has 5 targets which can appear and
disappear over time. The targets are evolved with constant linear velocity with
additive zero-mean noise. Targets initial states and time of birth/death is given in
Table 1. The GMPHD filter parameters is set as (ps,t )= 0.99, pd,t = 0.98, with
number of clutters reported in each scan.
TABLE 1

Initial location(m) Initial velocity (m/sec) time of birth/death(sec)
(1333.3,6866.7) (0,-136.7) 0/70
(-4000,4000) (106.7,-106.7) 10/85
(-6135.3,1594.1) (137.6,12.9) 25/95
(-4000,5000) (120.0,-40.0) 35/110
(4000,4000) (-106.7,-106.7) 50/120
[064] The initial parameters of adaptive birth component estimation
algorithm is set as (Tbc ) = 40, δ = 4, Δ = 2, Tnb = 3,TV = 60. The proposed
algorithm performance is measured using Optimal Sub-Pattern Assignment
(OSPA) metric [3] with cut-off parameter c=500m and the exponent
parameter d=1 . More is OSPA metric value, worse is the performance of the
filter. To evaluate the performance of disclosed adaptive birth Gaussian component as compared to the existing prior birth known GMPHD filter consider sensor network with no fusion.
[065] Graphical representations of FIG. 4 compare the standard GMPHD verses the adaptive birth Gaussian component (depicted in FIGs. 4 through 8 as adaptive GMPHD) based method disclosed herein, with no communication exchange, which means assumption is sensors are not linked with

each other and have no collaboration for target estimation. FIG. 4a depicts a simulation setup consisting of 4 sensors (shown with pentagon symbol) with restricted rectangular sensing region or coverage area. Initial location of the targets are depicted by star symbol and corresponding tracks are shown with different types of line pattern) and X represents environment clutter. Tracked target trajectories are shown using hashed lines. With reference to FIG. 4a, FIG. 4b depicts target tracking using GMPHD (standard GMPHD) with target’s initial position but no inter-sensor information exchange. With reference to FIG. 4a, FIG. 4c depicts target tracking using the disclosed adaptive GMPHD with no initial targets and no inter-sensor information fusion with generation of new birth components (birth targets depicted as triangles). The FIG. 4 (d) represents the ground truth (gt) of target number present in the sensor sensing region and the estimation done by the GMPHD and the method( adaptive GMPHD), which utilizes adaptive birth Gaussian component estimation. FIG. 4e depicts optimal Sub-pattern Assignment (OSPA) performance metric comparing GMPHD with the disclosed adaptive GMPHD using the adaptive birth Gaussian component estimation. It is observed that with uninitiated or unknown targets the GMPHD filter fails to track whereas with the method disclosed, new birth component are identified when target enter the sensor sensing region..
[066] FIG. 5 shows individual sensors local target number estimation in collaborative GMPHD. The collaborative GMPHD, refers to information exchange between the sensors to arrive at consensus on number of targets using the method disclosed. Specifically, FIG. 5a, 5b, 5c, 5d depicts actual number of targets ( ground truth) shown by solid line, GMPHD estimate by dots and adaptive GMPHD estimate by dashed line for sensor S1, S2, S3, S4 respectively. FIG. 6 shows overall sensors target estimation for collaborative GMPHD. Fig 6a, 6b, 6c, 6d depicts overall actual number of targets (ground truth shown by solid line), GMPHD estimates (by dotted line) and disclosed GMPHD (by dashed line) for sensor S1, S2, S3, S4 respectively shows the estimated number of targets converges in MSMTT framework. The FIGs. 5 and 6 also show that the estimate by adaptive GMPHD has a delayed in response to identify target track as it

requires a batch of measurements. FIG. 7 shows OSPA performance for collaborative GMPHD. . The FIG. 7a depicts Sensor S1, the FIG. 7b depicts Sensor S2, the FIG. 7c depicts Sensor S3 and the FIG. 7d depicts Sensor S4 showing similar performance. It can be observed from FIG. 7a through 7d that the OSPA performance metric for both GMPHD and the method disclosed (adaptive GMPHD) is almost identical.
[067] FIG. 8 shows trajectories estimation by sensor 1 (S1) in both unknown and known birth prior terms in MS-MTT framework. No additional new birth component or birth targets (depicted by triangles) is generated along sensor perimeter in the FIG. 8a while transit from one sensor region to other. Over further investigation, it is observed that due to inter-sensor information exchange all the sensors are able to track multiple targets throughout as show in FIG. 8a and 8b and no new birth components are generated at the entry of new target from the neighbors sensing region. Tracked target trajectories are shown using hashed lines.
[068] The written description describes the subject matter herein to enable any person skilled in the art to make and use the embodiments. The scope of the subject matter embodiments is defined by the claims and may include other modifications that occur to those skilled in the art. Such other modifications are intended to be within the scope of the claims if they have similar elements that do not differ from the literal language of the claims or if they include equivalent elements with insubstantial differences from the literal language of the claims.
[069] It is to be understood that the scope of the protection is extended to such a program and in addition to a computer-readable means having a message therein; such computer-readable storage means contain program-code means for implementation of one or more steps of the method, when the program runs on a server or mobile device or any suitable programmable device. The hardware device can be any kind of device which can be programmed including e.g. any kind of computer like a server or a personal computer, or the like, or any combination thereof. The device may also include means which could be e.g. hardware means like e.g. an application-specific integrated circuit (ASIC), a

field-programmable gate array (FPGA), or a combination of hardware and software means, e.g. an ASIC and an FPGA, or at least one microprocessor and at least one memory with software modules located therein. Thus, the means can include both hardware means and software means. The method embodiments described herein could be implemented in hardware and software. The device may also include software means. Alternatively, the embodiments may be implemented on different hardware devices, e.g. using a plurality of CPUs.
[070] The embodiments herein can comprise hardware and software elements. The embodiments that are implemented in software include but are not limited to, firmware, resident software, microcode, etc. The functions performed by various modules described herein may be implemented in other modules or combinations of other modules. For the purposes of this description, a computer-usable or computer readable medium can be any apparatus that can comprise, store, communicate, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device.
[071] The illustrated steps are set out to explain the exemplary
embodiments shown, and it should be anticipated that ongoing technological
development will change the manner in which particular functions are performed.
These examples are presented herein for purposes of illustration, and not
limitation. Further, the boundaries of the functional building blocks have been
arbitrarily defined herein for the convenience of the description. Alternative
boundaries can be defined so long as the specified functions and relationships
thereof are appropriately performed. Alternatives (including equivalents,
extensions, variations, deviations, etc., of those described herein) will be apparent to persons skilled in the relevant art(s) based on the teachings contained herein. Such alternatives fall within the scope and spirit of the disclosed embodiments. Also, the words “comprising,” “having,” “containing,” and “including,” and other similar forms are intended to be equivalent in meaning and be open ended in that an item or items following any one of these words is not meant to be an exhaustive listing of such item or items, or meant to be limited to only the listed item or items. It must also be noted that as used herein and in the appended

claims, the singular forms “a,” “an,” and “the” include plural references unless the context clearly dictates otherwise.
[072] Furthermore, one or more computer-readable storage media may be utilized in implementing embodiments consistent with the present disclosure. A computer-readable storage medium refers to any type of physical memory on which information or data readable by a processor may be stored. Thus, a computer-readable storage medium may store instructions for execution by one or more processors, including instructions for causing the processor(s) to perform steps or stages consistent with the embodiments described herein. The term “computer-readable medium” should be understood to include tangible items and exclude carrier waves and transient signals, i.e., be non-transitory. Examples include random access memory (RAM), read-only memory (ROM), volatile memory, nonvolatile memory, hard drives, CD ROMs, DVDs, flash drives, disks, and any other known physical storage media.
[073] It is intended that the disclosure and examples be considered as exemplary only, with a true scope and spirit of disclosed embodiments being indicated by the following claims.

WE CLAIM:
1. A processor implemented method for Multi Target Tracking (MTT) in a Distributed Sensor (DS) network, the method comprising: at a sensor among a plurality of sensors in the DS network;
capturing, by a sensor data acquisition model implemented by the
processor, a set of targets lying within a Field of Vision (FoV) of the
sensor from multiple targets present in a total area covered by the
plurality of sensors (302), wherein the captured set of targets comprise
surviving targets represented by previous target state estimates ( Xit-1 )
computed for a preceding time instance and a plurality of birth targets spawned at a current time instant;
dynamically determining, by a birth target estimator implemented
by the processor, the plurality of birth targets from the captured set of
targets based on an adaptive birth Gaussian component estimation
approach (304), wherein each birth target among the plurality of birth
targets is associated with a mean ( my,t ) representing a vector having
information of current target state estimates and features observed at the
current time instant, a preset birth weight (ωy,t ) and a state co-variance;
obtaining, by a target estimator implemented by the processor,
local target state estimates for the sensor providing a set of local targets
by applying a Gaussian mixture based probability hypothesis density
(GMPHD) filter (306), wherein the GMPHD filter utilizes the mean
(my,t), the state covariance of each birth target and the previous target
state estimates (Xit-1 ), wherein the local target state estimates provides
local probability densities of local targets estimated by the sensor;
exchanging, by the target estimator implemented by the processor, with a set of neighboring sensors among the plurality of sensors, the local target state estimates and neighboring target state estimates of each neighboring sensor among a set of neighboring sensors (308); and

processing, by the target estimator implemented by the processor, the local target state estimates and the neighboring target state estimates over multiple iterations based on a Gaussian Mixture General Covariance Intersection (GM-GCI) pair wise fusion approach to arrive at a final target state estimates (310), wherein the final target estimates provides an overall probability density function of final targets in the DS network.
2. The method as claimed in claim 1, wherein dynamically determining the plurality of birth targets from the captured set of targets based on the adaptive birth Gaussian component estimation approach comprises:
recording, by the sensor data acquisition model implemented by the processor, target parameters corresponding to each target from the set of targets, wherein the target parameters are recorded as one of state dynamic variables and image features;
representing, by the sensor data acquisition model implemented by
the processor, the set of targets and the recorded target parameters
corresponding to each target as a current measurement set (Zit )
associated with the current time instant;
determining a set of probable birth elements (Zibc,t ) by discarding
measurements from the current measurement set (Zit ) , which have
measurement value below a first preset criteria when compared with the
previous target state estimates (Xit-1 );
generating an birth batch measurement set ( zgr,t ) for the set of
probable birth elements (Zibc,t ) from backward frame sequences of
size {detection uncertainity size (Δ) + birth measurement size (δ) ;
selecting birth elements from the number of measurements
recorded within the birth batch measurement set ( Zgr,t ) associated with
each of the birth element among the set of probable birth elements, wherein cardinality of the selected birth elements is greater than the birth measurement size ( δ ); and

determining the plurality of birth targets from the selected birth elements by iterating a plurality of steps for each of the selected birth element, wherein the steps comprise:
a) identifying a line model by connecting consecutive measurements in the birth batch measurement set ( Zgr,t ) in accordance with a second preset criteria;
b) adding the selected birth element to a birth cluster ( Zbt ) if a perpendicular distance (d ) of the measurement corresponding to the selected birth element from the line model is within a line model cluster threshold (
T ⊥ ) and satisfies the second preset criteria; and
c) identifying the added birth element as a birth target,
among the plurality of birth targets to be determined, if
the birth cluster (Zbt ) is greater than a birth target
surety minimum set size threshold (Tn ), wherein the
birth target is associated with the mean ( my,t ) and the
features observed at the current time instant, the preset
birth weight ( ω y,t ) and the state variance.
3. The method as claimed in claim 1, wherein obtaining the local target state estimates for the sensor providing the set of local targets comprises:
predicating local target state estimates based on the mean (my,t )
corresponding to each of the detected plurality of birth targets and the
previous target state estimates (Xit-1 ), wherein each of the predicted local
target state estimates is associated with an intensity value;
estimating a set of possible target tracks using the predicted local
target state estimates and the current measurement set (Zit );
updating intensity value of each of the predicted local target state estimates based on the estimated set of possible target tracks;

identifying and pruning weaker targets from the predicted local target state estimates, wherein updated intensity value of the identified and pruned weaker targets is below a predefined intensity threshold (T); and
merging remaining targets from the predicted target state estimates based on a distance criteria to obtain the local target state estimates for the sensor.
4. The method as claimed in claim 1, wherein processing the local target
state estimates and the neighboring target state estimates over multiple
iterations based on the GM-GCI pair wise fusion approach to arrive at
the final target estimates comprises:
fusing a local target with a neighbor target if the local target estimates is present in the received neighboring local target estimates of a neighboring sensor based on the GM-GCI pair wise fusion approach to obtain a set of fused targets in accordance with fusion parameters W that are intensities associated with the local probability densities of each sensor;
marking unfused local targets and unfused neighbor targets;
reaching an agreement of targets state estimates with neighboring sensor based on the set of fused targets and the marked unfused targets;
iterating the GM-GCI pair wise fusion approach for all the plurality of sensors at multiple levels of iteration; and
terminating the iterations at a level among the multiple levels when an agreement criteria is satisfied to provide the final target state estimates as the overall probability density function.
5. The method as claimed in claim 1, wherein the first preset criteria is a
distance threshold ( Tbc ) when the target parameters are recorded in terms

of the state dynamic variables and a minimum 12-norm of feature
when the target parameters are recorded in terms of the image features.
6. The method as claimed in claim 1, wherein the second preset criteria is a
maximum possible displacement ( T dis ) value set when the target
parameters are recorded in terms of the state dynamic variables and a
12-norm of feature displacement thresholding when the target parameters are recorded in terms of the image features.
7. The method as claimed in claim 1, wherein the state variance refers to a birth covariance (Py,t ) which depends on a state noise of the birth target when the target parameters are recorded in terms of the state dynamic variables and the state variance provides spread of bounding box and scale changes when the target parameters are recorded in terms of the image features.
8. A Distributed Sensor (DS) network (100) for Multi Target Tracking (MTT), the DS network (100) comprising:
a plurality of sensors, wherein each sensor in the DS network comprises:
a memory (202) storing instructions;
Input/Output (I/O) interface (206); and
a processor (204) coupled to the memory (202) via the one or
more I/O interfaces (206), wherein the processor (204) is
configured by the instructions to:
capture a set of targets lying within a Field of Vision (FoV) of the
sensor from multiple targets present in a total area covered by the
plurality of sensors, wherein the captured set of targets comprise
surviving targets represented by previous target state estimates ( Xit-1 )
computed for a preceding time instance and a plurality of birth targets spawned at a current time instant;

dynamically determine the plurality of birth targets from the
captured set of targets based on an adaptive birth Gaussian component
estimation approach, wherein each birth target among the plurality of
birth targets is associated with a mean (my,t ) representing a vector having
information of current target state estimates and features observed at the
current time instant, a preset birth weight (ωy,t ) and a state co-variance;
obtain local target state estimates for the sensor providing a set of
local targets by applying a Gaussian mixture based probability hypothesis
density (GMPHD) filter, wherein the GMPHD filter utilizes the
mean (my,t ), the state covariance of each birth target and the previous
target state estimates (Xit-1 ), wherein the local target state estimates
provides local probability densities of local targets estimated by the sensor;
exchange with a set of neighboring sensors among the plurality of sensors, the local target state estimates and neighboring target state estimates of each neighboring sensor among a set of neighboring sensors; and
process the local target state estimates and the neighboring target state estimates over multiple iterations based on a Gaussian Mixture General Covariance Intersection (GM-GCI) pair wise fusion approach to arrive at a final target state estimates, wherein the final target estimates provides an overall probability density function of final targets in the DS network.
9. The DS network (100) as claimed in claim 8, wherein the processor ( 204) is configured to dynamically determine the plurality of birth targets from the captured set of targets based on the adaptive birth Gaussian component estimation approach by:

recording target parameters corresponding to each target from the set of targets, wherein the target parameters are recorded as one of state dynamic variables and image features;
representing the set of targets and the recorded target parameters
corresponding to each target as a current measurement set (Zti ) associated
with the current time instant;
determining a set of probable birth elements (Zibc,t ) by discarding
measurements from the current measurement set (Zti ) , which have
measurement value below a first preset criteria when compared with the
previous target state estimates (Xit-1 );
generating an birth batch measurement set (Zgr,t ) for the set of
probable birth elements (Zibc,t ) from backward frame sequences of
size {detection uncertianity size (Δ)+birth measurement size(δ)} ;
selecting birth elements from the number of measurements
recorded within the birth batch measurement set (Z gr,t ) associated with
each of the birth element among the set of probable birth elements, wherein cardinality of the selected birth elements is greater than the birth measurement size (δ ); and
determining the plurality of birth targets from the selected birth elements by iterating a plurality of steps for each of the selected birth element, wherein the steps comprise:
a) identifying a line model by connecting consecutive measurements in the birth batch measurement set (Z gr,t) in accordance with a second preset criteria;
b) adding the selected birth element to a birth cluster (Zbt ) if a perpendicular distance (d ) of the measurement corresponding to the selected birth element from the line model is within a line model cluster threshold (
T ⊥) and satisfies the second preset criteria; and

c) identifying the added birth element as a birth target,
among the plurality of birth targets to be determined, if
the birth cluster (Zbt ) is greater than a birth target
surety minimum set size threshold ( Tnb ), wherein the
birth target is associated with the mean (m γτ ) and the
features observed at the current time instant, the preset
birth weight (ωγ,τ ) and the state variance.
10. The DS network (100) as claimed in claim 8, wherein the processor (204)
is configured to obtain the local target state estimates for the sensor
providing the set of local targets by:
prediciting local target state estimates based on the mean ( mγ,τ )
corresponding to each of the detected plurality of birth targets and the
previous target state estimates (Xti-1 ), wherein each of the predicted local
target state estimates is associated with an intensity value;
estimating a set of possible target tracks using the predicted local
target state estimates and the current measurement set ( Zit );
updating intensity value of each of the predicted local target state estimates based on the estimated set of possible target tracks;
identifying and pruning weaker targets from the predicted local target state estimates, wherein updated intensity value of the identified and pruned weaker targets is below a predefined intensity threshold (T); and
merging remaining targets from the predicted target state estimates based on a distance criteria to obtain the local target state estimates for the sensor.
11. The DS network (100) as claimed in claim 8, wherein processing the
local target state estimates and the neighboring target state estimates over

multiple iterations based on the GM-GCI pair wise fusion approach to arrive at the final target estimates comprises:
fusing a local target with a neighbor target if the local target estimates is present in the received neighboring local target estimates of a neighboring sensor based on the GM-GCI pair wise fusion approach to obtain a set of fused targets in accordance with fusion parameters W that are intensities associated with the local probability densities of each sensor;
marking unfused local targets and unfused neighbor targets;
reaching an agreement of targets state estimates with neighboring sensor based on the set of fused targets and the marked unfused targets;
iterating the GM-GCI pair wise fusion approach for all the plurality of sensors at multiple levels of iteration; and
terminating the iterations at a level among the multiple levels when an agreement criteria is satisfied to provide the final target state estimates as the overall probability density function.
12. The DS network (100) as claimed in claim 8, wherein the first preset
criteria is a distance threshold ( Tbc ) when the target parameters are
recorded in terms of the state dynamic variables and a minimum
of feature when the target parameters are recorded in terms of the image features.
13. The DS network (100) as claimed in claim 8, wherein the second preset
criteria is a maximum possible displacement ( Tdis ) value set when the
target parameters are recorded in terms of the state dynamic variables and
a l2-norm of feature displacement thresholding when the target
parameters are recorded in terms of the image features.

14. The DS network (100) as claimed in claim 8, wherein the state variance refers to a birth covariance (Pγ,r ) which depends on a state noise of the
birth target when the target parameters are recorded in terms of the state dynamic variables and the state variance provides spread of bounding box and scale changes when the target parameters are recorded in terms of the image features.