Description:TECHNICAL FIELD
[0001] The present disclosure relates, in general, to a radar system, and more specifically, relates to a system and method for machine learning-based automatic target recognition using doppler return signals.

BACKGROUND
[0002] Typically, the Doppler returns obtained from a Doppler Radar are analysed by a trained human operator to identify the presence of a possible upcoming threat in the surveillance zone. Incorporation of an automatic target classification module in the radar system can aid in reducing the workload of the human operators significantly and at the same time can analyse more data.
[0003] An example of an existing technique is recited in a patent G01S13/5246 entitled, “Doppler radar system for the classification of moving targets”. The patent describes a method to classify moving objects using transmissions of interrupted pulse-modulated waves based upon the phase/frequency shift resulting from the moving objects, with reference to the transmitted signals. The method focuses on the component of radar echoes outside the main Doppler response line generated by the instantaneous relative speed of the main reflection from the target that contains additional information in their fine frequency structure.
[0004] Another example is recited in a patent US7567203B2 entitled “Classification system for radar and sonar applications”. The patent describes a radar-based target classification system and method for aircraft surveillance. The classifier provides tracks with an updated probability value based on their likelihood to conform to aircraft and non-aircraft-specific behaviour. The track classifier identifies false tracks arising from weather and biological targets and can detect aircraft lacking Secondary Surveillance Radar data. Different features and combinations of features are evaluated using a clustering performance index and used to discriminate between aircraft and false tracks.
[0005] Yet another example is recited in a publication 3948/CHE/2011 entitled “Doppler Based Classifier for Automated Target Recognition” relates to a scheme for classification of targets from Doppler returns of the electromagnetic transceiver using Wavelet transform coefficients. Every 250ms of Doppler signals are collected and transformed into wavelet coefficients. The unique features of the method enable tracking of the time and frequency information in the input signal associated with a particular kind of target. The wavelet coefficients are then analysed to identify target-specific parameters, and the nature of the target is estimated. However, the existing techniques suffer from significant drawbacks such as an increase in the workload of human operators and analyse of fewer data.
[0006] Therefore, it is desired to overcome the drawbacks, shortcomings, and limitations associated with existing solutions, and develop an efficient system that analyses the Doppler return signal and generates the classification result automatically, thereby reducing the workload of human operators significantly and at the same time analyse more data.

OBJECTS OF THE PRESENT DISCLOSURE
[0007] An object of the present disclosure relates, to a radar system, and more specifically, relates to a system and method for machine learning-based automatic target recognition using doppler return signals.
[0008] Another object of the present disclosure is to provide a system that aids in reducing the workload of the human operator significantly.
[0009] Another object of the present disclosure is to provide a system that analyses more data.
[0010] Another object of the present disclosure is to provide a system that analyses the Doppler returns and generates the classification result automatically and made available on demand.
[0011] Another object of the present disclosure is to provide a system that uses two different machine learning techniques significantly that increase the recognition performance.
[0012] Yet another object of the present disclosure is to provide a system that ensures smooth and consistent output.

SUMMARY
[0013] The present disclosure relates to a radar system, and more specifically, relates to a system and method for machine learning-based automatic target recognition using doppler return signals. The main objective of the present disclosure is to overcome the drawback, limitations, and shortcomings of the existing system and solution, by providing a system for the classification of targets from Doppler return signals obtained from a radar employing machine learning (ML) techniques.
[0014] The present disclosure relates to a transmitter that emits a signal at a predetermined frequency. A receiver that detects a return signal corresponding to the transmitted signal, the return signal being modulated by a target. A processor operatively coupled to the transmitter and the receiver, the processor configured to collect, at a data collection unit, the return signals being modulated by the target at a predefined period. The data collection unit is coupled to a non-target predictor that converts the collected return signals to segments of a one-second sample. The non-target predictor is configured to count changes in return signals from positive to negative and from negative to positive, compute and add the magnitude square of the return signals and determines statistical pattern similarity of the return signals with a pool of non-target return signal.
[0015] The non-target predictor is coupled to an appropriate model selection unit that selects between a near target model or a far target model based on the nature of the return signals through a switch. The appropriate model selection unit is coupled to a first classifier unit that identifies the target in the return signals. The selection of the near target model and far target model is performed by combining signal-to-noise ratio, distance from the radar to target, and intensity of return signals.
[0016] Further, the first classifier unit determines the category of the target and assigns scores to each category of the target. The generation of confidence scores is performed using radial basis function kernel and Viterbi algorithm. The first classifier unit is configured to compute different frequencies associated with the return signals in Hertz and mel-frequency domains. In addition, the first classifier unit is coupled to a second classifier unit to re-evaluate the identified categories of the target. The second classifier unit is adapted to assign likelihood scores to each category of the target, wherein the likelihood score is converted to a ratio by dividing the difference value of the maximum two likelihood values. The return signals are analysed through a weighted combination of the first classifier unit and the second classifier unit for the identification of the target in the return signals, thereby generating the classification result automatically. The use of two different machine learning techniques significantly increases the recognition performance and aid in reducing the workload of the human operator significantly while at the same time analysing more data.
[0017] Finally, the second classifier unit is coupled to a decision-making unit to produce either faster or stable outputs based on the user priority of threats. Accordingly, the decision-making unit analyses a sequence of classified labels corresponding to the return signals to estimate the nature of the target. Besides, the decision-making unit is coupled with a display unit to facilitate smooth and consistent output in the display unit.
[0018] Various objects, features, aspects, and advantages of the inventive subject matter will become more apparent from the following detailed description of preferred embodiments, along with the accompanying drawing figures in which like numerals represent like components.

BRIEF DESCRIPTION OF THE DRAWINGS
[0019] The following drawings form part of the present specification and are included to further illustrate aspects of the present disclosure. The disclosure may be better understood by reference to the drawings in combination with the detailed description of the specific embodiments presented herein.
[0020] FIG. 1A illustrates an exemplary representation of a target classification system, in accordance with an embodiment of the present disclosure.
[0021] FIG. 1B illustrates an exemplary block diagram of the proposed ML-based automatic target recognition system, in accordance with an embodiment of the present disclosure.
[0022] FIG. 2 illustrates an exemplary flow chart of a method for automated classification of targets, in accordance with an embodiment of the present disclosure.

DETAILED DESCRIPTION
[0023] The following is a detailed description of embodiments of the disclosure depicted in the accompanying drawings. The embodiments are in such detail as to clearly communicate the disclosure. If the specification states a component or feature “may”, “can”, “could”, or “might” be included or have a characteristic, that particular component or feature is not required to be included or have the characteristic.
[0024] As used in the description herein and throughout the claims that follow, the meaning of “a,” “an,” and “the” includes plural reference unless the context clearly dictates otherwise. Also, as used in the description herein, the meaning of “in” includes “in” and “on” unless the context clearly dictates otherwise.
[0025] The present disclosure relates, to a radar system, and more specifically, relates to a system and method for machine learning-based automatic target recognition using doppler return signals. The proposed system disclosed in the present disclosure overcomes the drawbacks, shortcomings, and limitations associated with the conventional system by providing a radar system configured for the classification of targets from Doppler returns obtained from a radar employing machine learning (ML) techniques. The proposed scheme takes a segment of Doppler return into account and classifies it utilizing the trained ML models. The present disclosure can be described in enabling detail in the following examples, which may represent more than one embodiment of the present disclosure.
[0026] The present disclosure relates to the detection of targets from Doppler returns using ML techniques. The Doppler radar provides velocity data about a moving target at a distance in the form of Doppler returns, by bouncing an electromagnetic signal off a desired target. The Doppler returns are collected every one second and analysed through a weighted combination of two ML techniques for identification of candidate target in Doppler returns by a first type of machine learning framework, assigning a confidence score, and re-verifying the target identification by the second type of machine learning framework. The detection results are then passed to the final decision-making module, which takes a sequence of recognized labels containing the last five outputs as input, analyses them and the nature of the target is estimated and displayed in a display unit.
[0027] The advantages achieved by the system of the present disclosure can be clear from the embodiments provided herein. The system can aid in reducing the workload of the human operator significantly and at the same time analyse more data. The system analyses the Doppler returns and generates the classification result automatically and made available on demand. The use of two different machine learning techniques significantly increases the recognition performance. The description of terms and features related to the present disclosure shall be clear from the embodiments that are illustrated and described; however, the invention is not limited to these embodiments only. Numerous modifications, changes, variations, substitutions, and equivalents of the embodiments are possible within the scope of the present disclosure. Additionally, the invention can include other embodiments that are within the scope of the claims but are not described in detail with respect to the following description.
[0028] FIG. 1A illustrates an exemplary representation of a target classification system, in accordance with an embodiment of the present disclosure.
[0029] Referring to FIG. 1A, target classification system 100 (also referred to as a system 100, herein) configured for classification of targets from Doppler return signal obtained from a radar employing machine learning (ML) techniques. The proposed scheme takes a segment of Doppler return signals into account and classifies it utilizing the trained ML models. The system 100 can include a radar having a transmitter 102 and a receiver 104 connected to a processor 106. The processor 106 is operatively coupled to a learning engine 108 i.e., a machine learning engine. Further, the system 100 can include a data collection unit 110, a non-target predictor 112, an appropriate model selection unit 114, a first classifier unit 116, a second classifier unit 118, a decision-making unit 120 and radar display 122 shown in FIG. 1B.
[0030] In an embodiment, the transmitter 102 and receiver 104 are directed at the same general area in an environment. The processor 106 is operatively coupled to a learning engine 108 according to the present disclosure, as shown in FIG. 1A. The learning engine 108 can include a first classifier unit 116 and a second classifier unit 118. The transmitter 102 emits a signal, where the signal can be electromagnetic or acoustic. The frequency of the signal can be adjusted according to a particular application and the desired resolution. The signal can be in the form of a continuous tone or a pulse train. The signal is reflected by an object or target in the environment. The target can be moving, have moving parts or both e.g., vehicle, people, clutter, and the likes. The reflected signal is detected by the receiver 104.
[0031] In an embodiment, the processor 106 is operatively coupled to the data collection unit 110, where the data collection unit 110 is adapted to collect the return signals from the radar at a predefined period. The return signals are modulated by the target and are collected every one second and analysed through a weighted combination of two ML techniques for the identification of candidate targets in the return signals.
[0032] The data collection unit 110 coupled to the non-target predictor 112, the non-target predictor 112 receives the return signals and converts a continuous input to the segment of a one-second sample and discards the data if it satisfies certain conditions. The non-target predictor 112 adapted to count changes in return signals (also referred to as Doppler returns signal) from positive to negative and from negative to positive and less than 10 counts signal can be discarded. The non-target predictor 112 can compute and add the magnitude square of the return signals with an addition value of below 10 discarded. The non-target predictor 112 determines statistical pattern similarity of the return signals with the pool of ‘non-target’ Doppler returns using the machine learning technique.
[0033] The appropriate model selection unit 114 coupled to the non-target predictor 112 and adapted to select through a switch between a near target model and a far target model based on the nature of the return signals. The selection of the near target model and the far target model is performed by combining signal-to-noise ratio, distance from Doppler radar to target, and intensity of Doppler return signals.
[0034] The first classifier unit 116 coupled to the appropriate model selection unit 114. The first classifier unit 116 is adapted to identify the candidate target (also referred to as target, herein) in the return signals. The candidate target in the return signal can be identified by computing different frequencies associated with the Doppler return signals in Hertz and mel-frequency domains. The first classifier unit 116 determines the category of the candidate target and assigns scores to each category of the candidate target. The generation of confidence scores is performed using radial basis function kernel and Viterbi algorithm.
[0035] The second classifier unit 118 coupled to the first classifier unit 116 and adapted to
re-evaluate the identified categories of the candidate target and assign likelihood scores to each category of the candidate target. The use of two different machine learning techniques significantly increases the recognition performance. The log-likelihood confidence is converted to the ratio for the second machine learning technique by dividing the difference value of the maximum two likelihood values.
[0036] In an exemplary implementation, the Doppler return signals are classified in one or more categories of the candidate target as vehicle and person. System 100 declares vehicles if both types of machine learning frameworks i.e., the first classifier unit 116 and the second classifier unit 118 declare the return signal as a vehicle. If the first one declares as the vehicle with more than 80% probability, and the second one classifies as a person, then the final decision is taken as the vehicle. If the first one declares person, and the second one declares the vehicle, the second technique confidence is compared to a pre-defined threshold for vehicle classification.
[0037] Accordingly, system 100 declares persons if both techniques declare the person. If the first one declares as the vehicle with less than 80% probability, and the second one declares as the person, then it is declared as the person. In all the other combinations of two technique outputs, which are not mentioned above, it is declared as clutter.
[0038] Finally, the decision-making unit 120 is coupled to the second classifier unit 118 and configured to produce either faster or stable outputs based on the user priority of threats. The decision-making unit 120 by a rule-based method collects five past target classification results corresponding to the past five seconds of Doppler return signals and a single class label is estimated. The display unit 122 coupled with the decision-making unit 120 to provide smooth and consistent output to display the final target class.
[0039] In another exemplary implementation, the collection of five past target classification results corresponding to the past five seconds of Doppler returns includes a sequence of vehicle, person, and clutter. If the sequence contains only vehicle/person/clutter, it declares the respective class. System 100 declares the vehicle when the sequence contains two or more vehicles and rest clutters. For a sequence of two or more vehicles and rest persons, then the system declares it as the vehicle. If the sequence contains two or more persons and the rest clutter, then the system declares it as the person. The system declares the person when it contains three or more persons and the rest vehicles. In all other combinations of vehicle, person and clutter in the sequence, the system 100 declares it as clutter. If the sequence of the decision-making unit 120 has only a vehicle, then it outputs the vehicle for the next five seconds. If the sequence of the decision-making unit 120 has only the person, then it is forced to output the person for the next five seconds.
[0040] Thus, the present invention overcomes the drawbacks, shortcomings, and limitations associated with existing solutions, and provides an efficient system that aids in reducing the workload of the human operator significantly and at the same time analyses more data. The use of two different machine learning techniques generates the classification result automatically that significantly increases the recognition performance and is made available on demand. Further, the system ensures smooth and consistent output in the display unit.
[0041] FIG. 2 illustrates an exemplary flow chart of a method for classification of targets, in accordance with an embodiment of the present disclosure.
[0042] Referring to FIG. 2, method 200 includes block 202, the transmitter emits the signal at the predetermined frequency. At block 204, the receiver can detect a return signal corresponding to the transmitted signal. The return signal is modulated by the target. The processor is operatively coupled to the transmitter and receiver. At block 206, the data collection unit can collect the return signals being modulated by the target. At block 208, the non-target predictor can convert the collected return signals to segments of one-second samples.
[0043] At block 210, the appropriate model selection unit can select a near target model and a far target model based on the nature of the return signals through a switch. At block 212, a first classifier unit can identify the target in the return signals, the first classifier unit determines the category of the target and assigns scores to each category of the target. In block 214, the second classifier unit can re-evaluate the identified categories of the target and at block 216 analyse, by the decision-making unit, a sequence of classified labels corresponding to the return signals to estimate the nature of the target.
[0044] It will be apparent to those skilled in the art that the system 100 of the disclosure may be provided using some or all of the mentioned features and components without departing from the scope of the present disclosure. While various embodiments of the present disclosure have been illustrated and described herein, it will be clear that the disclosure is not limited to these embodiments only. Numerous modifications, changes, variations, substitutions, and equivalents will be apparent to those skilled in the art, without departing from the spirit and scope of the disclosure, as described in the claims.

ADVANTAGES OF THE PRESENT INVENTION
[0045] The present disclosure provides a system that aids in reducing the workload of the human operator significantly.
[0046] The present disclosure provides a system that analyses more data.
[0047] The present disclosure provides a system that analyses the Doppler returns and generates the classification result automatically and made available on demand.
[0048] The present disclosure provides a system that uses two different machine learning techniques that significantly increases the recognition performance.
[0049] The present disclosure provides a system that ensures smooth and consistent output.
, Claims:1. A system (100) for automated classification of targets, the system comprising:
a transmitter (102) that emits a signal at a predetermined frequency;
a receiver (104) detects a return signal corresponding to the transmitted signal, the return signal being modulated by a target;
a processor (106) operatively coupled to the transmitter (102) and the receiver (104), the processor configured to:
collect, at a data collection unit (110), the return signals being modulated by the target;
convert, at a non-target predictor (112), the collected return signals to segments of a one-second sample;
select, by an appropriate model selection unit (114), a near target model or a far target model based on the nature of the return signals through a switch;
identify, by a first classifier unit, the target in the return signals, the first classifier unit determines the category of the target and assigns scores to each category of the target;
re-evaluate, by a second classifier unit, the identified categories of the target; and
analyze, by a decision-making unit, a sequence of classified labels corresponding to the return signals to estimate the nature of the target.
2. The system as claimed in claim 1, wherein the data collection unit (110) collects the return signals at a predefined period.
3. The system as claimed in claim 1, wherein the data collection unit (110) is coupled to the non-target predictor (112), the non-target predictor (112) is configured to:
count changes in the return signals from positive to negative and from negative to positive;
compute and add the magnitude square of the return signals; and
determines statistical pattern similarity of the return signals with a pool of non-target return signals.
4. The system as claimed in claim 1, wherein the non-target predictor (112) is coupled to the appropriate model selection unit (114), the selection of near target model or far target model is performed by combining signal to noise ratio, distance from a radar to target, and intensity of the return signals.
5. The system as claimed in claim 1, wherein the first classifier unit (116) is coupled to the appropriate model selection unit (114), the first classifier unit (116) is configured to compute different frequencies associated with the return signals in Hertz and mel-frequency domains.
6. The system as claimed in claim 1, wherein the second classifier unit (118) coupled to the first classifier unit (116), the second classifier unit (118) adapted to assign likelihood scores to each category of the target, wherein the likelihood score is converted to the ratio by dividing the difference value of the maximum two likelihood values.
7. The system as claimed in claim 1, wherein the generation of the scores is performed using radial basis function kernel and Viterbi algorithm.
8. The system as claimed in claim 1, wherein the decision-making unit (120) is coupled to the second classifier unit (118), the decision-making unit (120) is configured to produce either faster or stable outputs based on the user priority of threats.
9. The system as claimed in claim 1, wherein the return signals are analyzed through a weighted combination of the first classifier unit (116) and the second classifier unit (118) for identification of the target in the return signals.
10. A method (200) for automated classification of targets, the method comprising:
emitting (202), by a transmitter, a signal at a predetermined frequency;
detecting (204), at a receiver, a return signal corresponding to the transmitted signal, the return signal being modulated by a target, a processor operatively coupled to the transmitter and receiver;
collecting (206), at a data collection unit, the return signals being modulated by the target;
converting (208), at a non-target predictor, the collected return signals to segments of a one-second sample;
selecting (210), by an appropriate model selection unit, a near target model and a far target model based on the nature of the return signals through a switch;
identifying (212), by a first classifier unit, the target in the return signals, the first classifier unit determines the category of the target and assigns scores to each category of the target;
re-evaluating (214), by a second classifier unit, the identified categories of the target; and
analyzing (216), by a decision-making unit, a sequence of classified labels corresponding to the return signals to estimate the nature of the target.