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 POINT CLOUD DATA BASED OBJECT
DETECTION
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 the field of object detection, and, more particular, to a method and system for point cloud data based object detection.
BACKGROUND
[002] Autonomous vehicles are being developed rapidly because of their utility in multiple field. Autonomous vehicles are capable of sensing the environment and move without human intervention. The environment may include various vehicles, pedestrians, objects, traffic lights and lanes. The sensing of the environment is performed by utilizing sensors. Conventionally, point cloud data of the environment is obtained by utilizing the sensors. The obtained point cloud data is processed further for detecting the objects in the environment.
[003] Conventional methods mostly prefer fusion of one or more sensor along with a LiDAR (Light Detection And Ranging) to detect point cloud data for achieving better accuracy. The LiDAR works by measuring the distance to a target by illuminating the target with pulsed laser light and measuring the reflected pulses with a sensor. However, the cost of implementation is more. Further, the machine learning based conventional methods utilizing only LiDAR point cloud data is less efficient since the learning time is more. Hence there is a challenge in obtaining accurate object detection in minimum time.
SUMMARY [004] 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. For example, in one embodiment, a method for point cloud data based object detection is provided. The method includes receiving, a plurality of raw point cloud data associated with a plurality of objects pertaining to a scene, from a sensor, wherein each of the plurality of objects are labelled and bounded by a bounding box. Further, the method includes generating, a 3D voxel grid for each of the

plurality of the raw point cloud data. Further, the method includes, a confidence map associated with each 3D voxel grid corresponding to the bounding box, by utilizing a 3D Convolutional Neural Network (CNN), wherein the 3D CNN comprises a plurality of hidden layers, wherein the step of computing includes computing a plurality of relationship data associated with each 3D voxel grid corresponding to the bounding box, generating a set of feature maps corresponding to each hidden layer of the 3D CNN by utilizing the plurality of relationship data and computing a confidence map corresponding to each hidden layer by utilizing the set of feature maps. Further, the method includes computing a confidence map loss based on the computed confidence map and a predetermined confidence map corresponding to each bounding box. Further, the method includes simultaneously computing a bounding box loss based on an intersection between a bounding box under prediction and a reference bounding box. Finally the method includes detecting an object out of the plurality of objects in the scene by iteratively calculating a weighted average until loss is less than a predefined threshold wherein the weighted average is based on the confidence map loss and the bounding box loss.
[005] In another aspect, a system for point cloud data based object detection is provided. The system includes, an autonomous vehicle, one or more sensors and a computing device, wherein the computing device includes, at least one memory comprising programmed instructions, at least one hardware processor operatively coupled to the at least one memory, wherein the at least one hardware processor is capable of executing the programmed instructions stored in the at least one memories and an object detection unit, wherein the object detection unit is configured to receive, a plurality of raw point cloud data associated with a plurality of objects pertaining to a scene, from a sensor, wherein each of the plurality of objects are labelled and bounded by a bounding box. Further, the object detection unit is configured to generate a 3D voxel grid for each of the plurality of the raw point cloud data. Further, the object detection unit is configured to compute a confidence map associated with each 3D voxel grid corresponding to the bounding box, by utilizing a 3D Convolutional Neural

Network (CNN), wherein the 3D CNN comprises a plurality of hidden layers. The step of computing includes computing a plurality of relationship data associated with each 3D voxel grid corresponding to the bounding box, generating a set of feature maps corresponding to each hidden layer of the 3D CNN by utilizing the plurality of relationship data, and computing a confidence map corresponding to each hidden layer by utilizing the set of feature maps. Further, the object detection unit is configured to compute a confidence map loss based on the computed confidence map and a predetermined confidence map corresponding to each bounding box. Furthermore, the object detection unit is configured to simultaneously compute a bounding box loss based on an intersection between a bounding box under prediction and a reference bounding box. Finally, the object detection unit is configured to detect an object out of the plurality of objects in the scene by iteratively calculating a weighted average until loss is less than a predefined threshold wherein the weighted average is based on the confidence map loss and the bounding box loss.
[006] In yet another aspect, a computer program product comprising a non-transitory computer-readable medium having the object detection unit is configured to embodied therein a computer program for method and system for point cloud data based object detection is provided. The computer readable program, when executed on a computing device, causes the computing device to receive a plurality of raw point cloud data associated with a plurality of objects pertaining to a scene, from a sensor, wherein each of the plurality of objects are labelled and bounded by a bounding box. Further, the computer readable program, when executed on a computing device, causes the computing device to generate a 3D voxel grid for each of the plurality of the raw point cloud data. Further, the computer readable program, when executed on a computing device, causes the computing device to compute a confidence map associated with each 3D voxel grid corresponding to the bounding box, by utilizing a 3D Convolutional Neural Network (CNN), wherein the 3D CNN comprises a plurality of hidden layers. The step of computing includes computing a plurality of relationship data associated with each 3D voxel grid corresponding to the

bounding box. generating a set of feature maps corresponding to each hidden layer of the 3D CNN by utilizing the plurality of relationship data and computing a confidence map corresponding to each hidden layer by utilizing the set of feature maps. Further, the computer readable program, when executed on a computing device, causes the computing device to compute a confidence map loss based on the computed confidence map and a predetermined confidence map corresponding to each bounding box. Furthermore, the computer readable program, when executed on a computing device, causes the computing device to simultaneously compute a bounding box loss based on an intersection between a bounding box under prediction and a reference bounding box. Finally, the computer readable program, when executed on a computing device, causes the computing device to detect an object out of the plurality of objects in the scene by iteratively calculating a weighted average until loss is less than a predefined threshold wherein the weighted average is based on the confidence map loss and the bounding box loss.
[007] 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 [008] The accompanying drawings, which are incorporated in and constitute a part of this disclosure, illustrate exemplary embodiments and, together with the description, serve to explain the disclosed principles:
[009] FIG. 1 illustrates an exemplary networking environment for implementing a system for point cloud data based object detection.
[010] FIG. 2 is a functional block diagram of a system for point cloud data based object detection, according to some embodiments of the present disclosure.
[011] FIG. 3 illustrates an example pre-trained 3 Dimensional Convolution Neural Network (3D CNN) for point cloud data based object detection, in accordance with some embodiments of the present disclosure.

[012] FIG. 4 is an exemplary flow diagram for a processor implemented method for point cloud data based object detection, according to some embodiments of the present disclosure.
DETAILED DESCRIPTION OF EMBODIMENTS [013] 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.
[014] Embodiments herein provide a method and system for point cloud data based object detection. The system for point cloud data based object detection provides an accurate object detection by utilizing Convolutional Neural Network (CNN). Here a 5 layered 3D CNN with custom activation function is utilized. The custom activation function utilizes ReLU (Rectified Linear Unit) and sigmoid functions in same layer of the 3D CNN. Here, frames are converted into grids and a plurality of relationship data is calculated for each grid. The plurality of relationship data is further used for localization and accurate depth measurement of distance to objects. Further, a linear regression based custom loss function is utilized for accurate detection of the object. An implementation of the method and system for point cloud data based object detection is described further in detail with reference to FIGS. 1 through 4.
[015] Referring now to the drawings, and more particularly to FIG. 1 through 4, 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.
[016] FIG. 1 illustrates an exemplary networking environment for implementing a system for point cloud data based object detection, according to an example embodiment of the present subject matter. The system 100 for point cloud data based object detection, hereinafter referred to as the system 100, includes a computing device 104, an autonomous vehicle 102, a plurality if sensors 106 and a network 108. Here the computing device 104 can be one among a variety of computing systems, such as a laptop computer, a desktop computer, a notebook, a workstation, a cloud-based computing device, a router, a network gateway, a sensor gateway, a Wi-Fi access point and the like. In one implementation, the system 100 may be implemented in a cloud-based environment. In another implementation, the system 100 can be implemented in a cloud-edge environment and in yet another implementation, the system 100 can be implemented in a cloud-fog environment. The autonomous vehicle 102, one or more sensors 106 and the computing device 104 are communicatively coupled through a network 108. In an embodiment, the autonomous vehicle 102 includes a robot or a driverless car.
[017] In an embodiment, the network 108 may be a wireless or a wired network, or a combination thereof. In an example, the network 108 can be implemented as a computer network, as one of the different types of networks, such as virtual private network (VPN), intranet, local area network (LAN), wide area network (WAN), the internet, and such. The network 108 may either be a dedicated network or a shared network, which represents an association of the different types of networks that use a variety of protocols, for example, Hypertext Transfer Protocol (HTTP), Transmission Control Protocol/Internet Protocol (TCP/IP), and Wireless Application Protocol (WAP), to communicate with each other. Further, the network 108 may include a variety of network devices, including routers, bridges, servers, computing devices, storage devices. The network devices within the network 108 may interact with the each other through communication links.

[018] FIG. 2 illustrates a functional block diagram of the computing device 104 for point cloud data based object detection, according to some embodiments of the present disclosure. The computing device 104 includes or is otherwise in communication with one or more hardware processors, such as a processors 202, at least one memory such as a memory 204, an I/O interface 222. The memory 204 may include an object detection unit 220. The processor 202, memory 204, and the I/O interface 222 may be coupled by a system bus such as a system bus 208 or a similar mechanism.
[019] The I/O interface 222 may include a variety of software and hardware interfaces, for example, a web interface, a graphical user interface, and the like. The I/O interface 222 may include a variety of software and hardware interfaces, for example, interfaces for peripheral device(s), such as a keyboard, a mouse, an external memory, the one or more sensor devices, a printer and the like. Further, the I/O interface 222 may enable the computing device 104 to communicate with other devices, such as web servers and external databases.
[020] The I/O interface 222 can facilitate multiple communications within a wide variety of networks and protocol types, including wired networks, for example, local area network (LAN), cable, etc., and wireless networks, such as Wireless LAN (WLAN), cellular, or satellite. For the purpose, the I/O interface 222 may include one or more ports for connecting a number of computing systems with one another or to another server computer. The I/O interface 222 may include one or more ports for connecting a number of devices to one another or to another server.
[021] The one or more hardware processors 202 may 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 one or more hardware processors 202 is configured to fetch and execute computer-readable instructions stored in the memory 204.
[022] The memory 204 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, the memory 204 includes a plurality of modules 206 and a repository 210 for storing data processed, received, and generated by one or more of the modules 106 and the object detection unit 220. In an embodiment, the object detection unit 220 includes a voxel grid construction module (not shown in FIG. 2), confidence map computation module (not shown in FIG. 2), confidence map loss computation module (not shown in FIG. 2), bounding box loss computation module (not shown in FIG. 2) and a detecting module (not shown in FIG. 2). The modules 106 may include routines, programs, objects, components, data structures, and so on, which perform particular tasks or implement particular abstract data types.
[023] The memory 204 also includes module(s) 206 and a data repository 210. The module(s) 206 include programs or coded instructions that supplement applications or functions performed by the computing device 104 for point cloud data based object detection. The modules 106, amongst other things, can include routines, programs, objects, components, and data structures, which perform particular tasks or implement particular abstract data types. The modules 206 may also be used as, signal processor(s), state machine(s), logic circuitries, and/or any other device or component that manipulates signals based on operational instructions. Further, the modules 206 can be used by hardware, by computer-readable instructions executed by a processing unit, or by a combination thereof. The modules 206 can include various sub-modules (not shown). The modules 206 may include computer-readable instructions that supplement applications or functions performed by the computing device 104 for point cloud data based object detection.
[024] The data repository 210 may include image data obtained from the one or more sensors, voxel grid data and other data. Further, the other data amongst other things, may serve as a repository for storing data that is processed, received, or generated as a result of the execution of one or more modules in the

module(s) 206 and the modules associated with the point cloud data based object detection unit 220.
[025] Although the data repository 210 is shown internal to the computing device 104, it will be noted that, in alternate embodiments, the data repository 210 can also be implemented external to the computing device 200, where the data repository 210 may be stored within a database (not shown in FIG. 2) communicatively coupled to the computing device 104. The data contained within such external database may be periodically updated. For example, new data may be added into the database (not shown in FIG. 2) and/or existing data may be modified and/or non-useful data may be deleted from the database (not shown in FIG. 2). In one example, the data may be stored in an external system, such as a Lightweight Directory Access Protocol (LDAP) directory and a Relational Database Management System (RDBMS). In another embodiment, the data stored in the data repository 210 may be distributed between the computing device 204 and the external database (not shown).
[026] FIG. 3 illustrates an example of a pre-trained 3 Dimensional Convolution Neural Network (3D CNN) machine learning architecture 200 of the method for point cloud data based object detection, in accordance with some embodiments of the present disclosure. Now, referring to FIG. 2, the 4D voxel grid data 302 is provided as input to 3D CNN machine learning architecture. In an embodiment, the 3D CNN machine learning architecture includes 5 layers with varying filter size, filter count and stride. For example, the first layer 304 includes a convolution 3D filter with size (3x3x1), a filter count of 9 and a stride of (2, 2, 2). The second layer 306 includes a convolution 3D filter with size (3x3x3x9), a filter count of 19 and a stride of (2, 2, 2). The third layer 308 includes a convolution 3D filter with size (3x3x1x9), a filter count of 29 and a stride of (2, 2, 1). The fourth layer 310 includes a convolution 3D filter with size (3x3x2x9), a filter count of 39 and a stride of (2, 2, 1). The fifth layer 312 includes a convolution 3D filter with size (1x1x1x3x9 ), a filter count of 11 and a stride of (1,1,1). In an embodiment, the 3D CNN architecture of the present disclosure includes a custom activation unit 314. The custom activation unit 314

includes a combination of ReLU and sigmoid activation function. The bounding box extractor 316 extracts data associated with each bounding box among a plurality of bounding boxes.
[027] The object detection unit 220 of the computing device 104 can be configured to receive a plurality of raw 3D point cloud data associated with a plurality of objects pertaining to a scene from a sensor 106wherein each of the plurality of objects are labelled and each of the plurality of objects is bounded by the bounding box. In an embodiment, 3D Point Cloud Data (PCD) is a digital representation of volumetric information, most commonly portraying outdoor scenes or landscapes. The point cloud is a set of data points in space and are generally produced by 3D scanners.
[028] In an embodiment, the sensor can be a Velodyne HDL-64E sensor. The Velodyne LiDAR (Light Detection And Ranging) sensors are portable and can be fixed on mobile robots and vehicles. The Velodyne HDL-64E utilizes 64 laser beams spread over a 26.8° vertical angle and achieves a 360° horizontal field of view by spinning at 300-900 RPM around the base. The sensor can map up to 2.2 million points per second. The sensor includes a rotating head with a sensor array on one side mounted behind a glass panel. The laser arrays and photo sensors are used to scan the surroundings of the sensor are split into an upper level and a lower level, each including three lenses.
[029] Further, the object detection unit 220 of the computing device 104 can be configured to generate a 3D voxel grid for each of a plurality of the raw point cloud data to obtain a plurality of voxel grids. The voxel grids are a set of tiny boxes in space over the input point cloud data. The voxel grid forms a 3D box around each point and can be utilized as a fixed spatial resolution in the raw point cloud data. In an embodiment, the voxel grids are 3D arrays of size given by the range where the detections are required. In an embodiment, the voxel size can be the range between (0, 48), (-24, 24), (-1.65, 4.5).
[030] Further, the object detection unit 220 of the computing device 104 can be configured to compute a confidence map associated with each voxel grid corresponding to the bounding box by utilizing the 3D Convolutional Neural

Network. The method of computing the confidence map includes the following steps: (1) computing a plurality of relationship data associated with each voxel grid corresponding to the bounding box (2) generating a set of feature maps corresponding to each hidden layer of the 3D CNN by utilizing the plurality of relationship data and (3) computing a confidence map corresponding to each hidden layer by utilizing the set of feature maps. The set of feature maps includes a plurality of features including a shape of the object, a size of object and an intensity of object.
[031] In an embodiment a 5 layered with 3D convolutional neural network is utilized to detect an object in a scene. Each layer has a different number of filters. As the depth increases the number of filters increases. The increase in number of filters enables the identification of smallest distinguishable features from the input for detection of an object in the scene based on 3D point cloud data.
[032] In an embodiment, the first layer of the 3D convolutional network includes 9 filters with size (3x3x3x1) and the stride size is (2,2,2). The neurons utilizes a sigmoid function as the activation function. Here the filter number is chosen less so that the features extracted will be more prominent or are comparatively easy to detect. The first layer receives the voxel grids of the raw point cloud data.
[033] In an embodiment, the Layers 2, 3 and 4 are hidden Layers. The hidden layers have respectively 19, 29 and 39 filters of sizes(3x3x3x9), (3x3x3x19), (3x3x3x29) respectively. Here the number of filters are more to obtain abstract features from the input. The stride for the first hidden layer is of size(2,2,2). The sigmoid activation function is utilized in the first layer. The second hidden layer has a stride size of (2,2,1), where the step size of convolution along the 3rd dimension (z in point cloud) is 1. The number of filters in the second layer is 29. Sigmoid function is used as activation function in the second layer. The third hidden layer is same as the 2nd hidden layer but the number of filters is 39 and is extracting very abstract features for learning. The output layer

has 11 filters of size (3x3x3x39) and stride of size (1,1,1). The custom activation function and a custom loss function are utilized in the output layer.
[034] In an embodiment, the custom activation function ?? includes ReLU and sigmoid activation function. Here, the custom activation function utilizes the advantages of the ReLU and sigmoid functions. The ReLU function is a prominently utilized activation function in deep neural networks. The most important feature of ReLU is speed in the training process due to the simple computation of gradient for the ReLU. The gradient function for the ReLU is either 0 or 1. The mathematical representation of the ReLU is given in equation 1.
ƒ ( x ) = max (0,x)……………… (1)
In tangent networks, the gradients are smaller than the positive portion of ReLU.
The positive parts of ReLU are updated very rapidly as the training progresses. In
an embodiment, though ReLU is faster, ReLU may have negative bias value. In
such cases ReLU will reach a state called “dead neuron” where the output will
always be a zero, irrespective of the input. Hence, there is a challenge in utilizing
ReLU with all categories of data. To overcome the limitation of the ReLU, the
present disclosure utilizes two distinct activations in the same layer, i.e. output
layer. The two distinct activation function includes a sigmoid activation function
and ReLU activation function. A sigmoid activation function has a differentiable
function when compared to ReLU. The slope of the curve could be found on any
point and may not become completely zero at any point. The gradient is always a
non-zero value and the training speed is less compared to the ReLU. However,
utilizing both the sigmoid and the ReLU together in the output layer enhances
time as well as accuracy.
[035] Further, the object detection unit 220 of the computing device 104 can be configured to compute a confidence map loss based on the computed confidence map and a predetermined confidence map corresponding to each bounding box. The confidence map is obtained by concatenating the set of feature maps associated with each hidden layer of the 3D CNN.
[036] Further, the object detection unit 220 of the computing device 104 can be configured to simultaneously compute a bounding box loss based on an

intersection between a bounding box under prediction and a reference bounding box.
[037] In an embodiment, the prediction of bounding boxes is carried out utilizing the following steps:
Step 1. If a point is part of a bounding box.
Step 2: X being area of the predicted bounding box and being the area of ground truth box.
Step 3: Ih and Iw are the height and width of the intersection area,
then, I = Ih * Iw .
Step 4: U = X + Xg -I
Step 5: IoU = I/U
Step 6: Boundingbox loss = -In ( IoU ) In an embodiment, the IoU is taken to be a random variable sampled from Bernoulli distribution and the cross entropy variable is as given in equation 2. -p* In(IoU) - (1-p) ……………………… (2)
Here, In (1-IoU) = -In(IoU ). The bounding box is considered as a
single unit and the final layer is having a larger field than the previous layers. Hence the bounding box is localized in the final layer with the concatenated confidences from previous layers
[038] Further, the object detection unit 220 of the computing device 104 can be configured to detect, an object in the scene by iteratively calculating a weighted average until loss is minimum, wherein the weighted average is based on the confidence map loss and the bounding box loss. The weighted average (total_loss) of the confidence map loss and the bounding box loss is given in equation 3.

Total_loss = (0.4*confidence map loss +0.6*bounding box loss) ......

(3)
[039] In an embodiment, a method for the bounding box prediction is utilized, wherein the regression takes place with the bounding box as a whole unit. Unlike the other methods where the 4 variables in the bounding box are

treated as independent variables, the present disclosure is oversimplifying a good observation. So to make regression more useful, the present disclosure utilizes the said method for efficient localization.
[040] In an embodiment, the algorithm for calculating the custom loss function is as follows:
a. Input x: Label Data
b. Predicted x: Bounding Box Predicted
c. Output: Localization Error
d. for each point in the cloud():

[041] FIG. 4 is an exemplary flow diagram for a processor implemented method for point cloud data based object detection, according to some embodiments of the present disclosure. The method 400 may be described in the general context of computer executable instructions. Generally, computer executable instructions can include routines, programs, objects, components, data structures, procedures, modules, functions, etc., that perform particular functions or implement particular abstract data types. The method 400 may also be practiced in a distributed computing environment where functions are performed by remote processing devices that are linked through a communication network. The order in which the method 400 is described is not intended to be construed as a limitation, and any number of the described method blocks can be combined in any order to implement the method 400, or an alternative method. Furthermore,

the method 400 can be implemented in any suitable hardware, software, firmware, or combination thereof.
[042] At 402, the computing device 104, receives, by a one or more hardware processors, a plurality of raw point cloud data associated with a plurality of objects pertaining to a scene from a sensor, wherein each of the plurality of objects are labelled and each of the plurality of objects is bounded by a bounding box.
[043] At 404, the computing device 104generates, by the one or more hardware processors, a 3D voxel grid for each of the plurality of the raw point cloud data to obtain a plurality of voxel grids.
[044] At 406, the computing device 104computes, by the one or more hardware processors, a confidence map associated with each voxel .grid corresponding to the bounding box by utilizing a 3D Convolutional Neural Network, wherein the 3D CNN comprises a plurality of hidden layers, wherein the step of computing includes (i) computing a plurality of relationship data associated with each voxel grid corresponding to the bounding box (ii) generating a set of feature maps corresponding to each hidden layer of the 3D CNN by utilizing the plurality of relationship data and (iii) computing a confidence map corresponding to each hidden layer by utilizing the set of feature maps.
[045] At 408, the computing device 104, computes by the one or more hardware processors, a confidence map loss based on the computed confidence map and a predetermined confidence map corresponding to each bounding box. The confidence map is obtained by concatenating the set of feature maps associated with each hidden layer of the 3D CNN.
[046] At 410, the computing device 104simultaneously computes, by the one or more hardware processors, a bounding box loss based on an intersection between a bounding box under prediction and a reference bounding box.
[047] At 412, the computing device 104detects, by the one or more hardware processors, an object from a plurality of objects in the scene by iteratively calculating a weighted average until loss is less than a predefined

threshold, wherein the weighted average is based on the confidence map loss and the bounding box loss.
[048] 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.
[049] The embodiments of present disclosure herein addresses unresolved problem of accurate and fast object detection using point cloud data alone. Here, a combination of ReLU and sigmoid activation function are utilized in the output layer of the 3D CNN to reduce training time. Further, the system 100 considers the individual bounding box prediction based on bounding box loss and a confidence map loss. The weighted average function of the system 100 provides the accurate object detection.
[050] 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.
[051] 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.
[052] 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.
[053] 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. 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.
[054] 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, the method comprising:
receiving, by a one or more hardware processors, a plurality of raw point cloud data associated with a plurality of objects pertaining to a scene, from a sensor, wherein each of the plurality of objects are labelled and bounded by a bounding box (402);
generating, by the one or more hardware processors, a 3D voxel grid for each of the plurality of the raw point cloud data (404);
computing, by the one or more hardware processors, a confidence map associated with each 3D voxel grid corresponding to the bounding box, by utilizing a 3D Convolutional Neural Network (CNN), wherein the 3D CNN comprises a plurality of hidden layers, wherein the step of computing (406) comprising:
computing a plurality of relationship data associated with each 3D voxel grid corresponding to the bounding box,
generating a set of feature maps corresponding to each hidden layer of the 3D CNN by utilizing the plurality of relationship data, and
computing a confidence map corresponding to each hidden layer by utilizing the set of feature maps; computing, by the one or more hardware processors, a confidence map loss based on the computed confidence map and a predetermined confidence map corresponding to each bounding box (408);
simultaneously computing, by the one or more hardware processors, a bounding box loss based on an intersection between a bounding box under prediction and a reference bounding box (410); and
detecting, by the one or more hardware processors, an object out of the plurality of objects in the scene by iteratively calculating a weighted average until loss is less than a predefined threshold wherein the weighted average is based on the confidence map loss and the bounding box loss (412).

2. The method as claimed in claim 1, wherein an output layer of the 3D CNN is activated by a combination of ReLU (Rectified Linear Unit) and sigmoid function.
3. The method as claimed in claim 1, wherein the confidence map is obtained by concatenating the set of feature maps associated with each hidden layer of the 3D CNN.
4. The method as claimed in claim 1, wherein the 3D Convolutional Neural Network is pre-trained.
5. A system (100), the system (100) comprising:
at least one memory (204) storing programmed instructions;
one or more hardware processors (202) operatively coupled to the at least
one memory, wherein the one or more hardware processors (202) are capable
of executing the programmed instructions stored in the at least one memory
(204); and
an object detection unit (220), wherein the point cloud data based object
detection unit (120) is configured to:
receive a plurality of raw point cloud data associated with a plurality of
objects pertaining to a scene, from a sensor, wherein each of the plurality of
objects are labelled and bounded by a bounding box;
generate a 3D voxel grid for each of the plurality of the raw point cloud data;
compute a confidence map associated with each 3D voxel grid corresponding
to the bounding box, by utilizing a 3D Convolutional Neural Network (CNN),
wherein the 3D CNN comprises a plurality of hidden layers, wherein the step
of computing comprising:
computing a plurality of relationship data associated with each 3D voxel grid
corresponding to the bounding box,
generating a set of feature maps corresponding to each hidden layer of the 3D
CNN by utilizing the plurality of relationship data, and

computing a confidence map corresponding to each hidden layer by utilizing the set of feature maps;
compute a confidence map loss based on the computed confidence map and a predetermined confidence map corresponding to each bounding box; simultaneously compute a bounding box loss based on an intersection between a bounding box under prediction and a reference bounding box; and detect an object out of the plurality of objects in the scene by iteratively calculating a weighted average until loss is less than a predefined threshold wherein the weighted average is based on the confidence map loss and the bounding box loss.
6. The system of claim 5, wherein an output layer of the 3D CNN is activated by a combination of ReLU (Rectified Linear Unit) and sigmoid function.
7. The system of claim 5, wherein the confidence map is obtained by concatenating the set of feature maps associated with each hidden layer of the 3D CNN.
8. The system of claim 5, wherein the 3D Convolutional Neural Network is pre-trained.