System and method for Doppler coding for radar object detection

Abstract

A method of performing Doppler spectrum coding for radar object detection. A method includes receiving a radar tensor, the radar tensor including a range and angle combined Doppler. A coding vector is generated for each of the range and angle combinations. For each of the range and angle combinations, the encoding vector includes a predetermined number of reflection intensities of a maximum reflection intensity peak, a frequency of the reflection intensities, and a width of the reflection intensity peak. A set of encoded vector data is generated from the encoded vectors. Feature extraction is performed on the encoded vector data set using a feature extraction neural network to generate an extracted feature data set with extracted feature information for each of the range and angle combinations. Object detection is performed on the extracted feature data set using an object detection neural network to generate an object detection output for identifying the object.

Description

[0001] Introduction

[0002] The present disclosure relates to a system and method for Doppler spectrum encoding for radar for object detection using an object detection neural network.

[0003] Automotive radar is one of the most promising and fastest-growing civilian applications of radar technology. Vehicle-mounted radar provides a key enabling technology for the autonomous driving revolution, with the potential to improve everyone's daily life. Automotive radar, together with other sensors such as lidar (which stands for "light detection and ranging"), ultrasonic, and cameras, forms the backbone of self-driving cars and advanced driver assistance systems (ADASs). These technological advancements are enabled by extremely complex systems with long signal processing paths from the radar / sensor to the controller. Automotive radar systems are responsible for detecting objects and obstacles, their positions, and their speeds relative to the vehicle. Summary of the Invention

[0004] Disclosed herein is a method for performing Doppler spectrum encoding for radar object detection. The method includes receiving a radar tensor that includes Doppler for range and angle combinations. Generating, for each of the range and angle combinations, an encoding vector from the radar tensor. For each of the range and angle combinations, the encoding vector includes the reflection intensities of a predetermined number of maximum reflection intensity peaks, the frequencies of the reflection intensities, and the widths of the reflection intensity peaks. Generating, for each of the range and angle combinations, an encoding vector data set from the encoding vector. Performing feature extraction on the encoding vector data set using a feature extraction neural network to generate an extracted feature data set having extracted feature information for each of the range and angle combinations. Performing object detection on the extracted feature data set using an object detection neural network to generate an object detection output for identifying objects from the radar tensor.

[0005] Another aspect of the present disclosure may be that the predetermined number of reflection intensity peaks includes at least three and no more than five reflection intensity peaks.

[0006] Another aspect of the present disclosure may be that the encoding vectors are sorted in descending order of reflection intensity.

[0007] Another aspect of the present disclosure may be that the predetermined number of maximum reflection intensity peaks is based on the angle of the radar system used to capture the radar tensor.

[0008] Another aspect of the present disclosure may include generating, for each of the range and angle combinations, a frequency data set having the frequencies of the reflection intensities of each of the predetermined number of maximum reflection intensity peaks.

[0009] Another aspect of the present disclosure may be where a frequency data set is concatenated to the extracted feature data set to generate a concatenated data set.

[0010] Another aspect of the present disclosure may be where performing object detection on the extracted feature data set includes performing object detection on a concatenated data set including both the extracted feature data set and the frequency data set.

[0011] Another aspect of the present disclosure may include determining the accuracy of an object detection neural network by comparing the object detection output with a ground truth data set of radar tensors using a loss function.

[0012] Another aspect of the present disclosure may include adapting the parameters of an object detection neural network based on the loss function to generate an updated object detection neural network.

[0013] Another aspect of the present disclosure may include adapting the parameters of a feature extraction neural network based on the loss function to generate an updated object feature extraction neural network.

[0014] Disclosed herein is a non-transitory computer-readable storage medium containing programming instructions that, when executed by a processor, are operable to perform a method. The method includes receiving a radar tensor that includes Doppler for range and angle combinations. Generating, for each of the range and angle combinations, an encoded vector from the radar tensor. For each of the range and angle combinations, the encoded vector includes the reflection intensity of a predetermined number of maximum reflection intensity peaks, the frequency of the reflection intensity, and the width of the reflection intensity peaks. Generating an encoded vector data set from the encoded vectors for each of the range and angle combinations. Performing feature extraction on the encoded vector data set using a feature extraction neural network to generate, for each of the range and angle combinations, an extracted feature data set having extracted feature information. Performing object detection on the extracted feature data set using an object detection neural network to generate an object detection output for identifying an object from the radar tensor.

[0015] This disclosure relates to a vehicle system. The system includes at least one radar sensor and a controller in communication with the at least one radar sensor. The at least one radar sensor is configured to capture information for generating a radar tensor. The controller is configured to receive the radar tensor including Doppler of range and angle combinations, and generate an encoded vector for each of the range and angle combinations from the radar tensor. For each of the range and angle combinations, the encoded vector includes the reflection intensities of a predetermined number of maximum reflection intensity peaks, the frequencies of the reflection intensities, and the widths of the reflection intensity peaks. The controller is further configured to generate an encoded vector data set from the encoded vector for each of the range and angle combinations, and perform feature extraction on the encoded vector data set using a feature extraction neural network. For each of the range and angle combinations, the feature extraction neural network generates an extracted feature data set with the extracted feature information. The controller is also configured to perform object detection on the extracted feature data set using an object detection neural network to generate an object detection output for identifying an object from the radar tensor.

[0016] The present invention further includes the following solutions.

[0017] Solution 1. A method for performing Doppler spectrum encoding for radar object detection, the method comprising:

[0018] Receiving a radar tensor, wherein the radar tensor includes Doppler for a plurality of range and angle combinations;

[0019] Generating an encoded vector for each of the plurality of range and angle combinations from the radar tensor, wherein for each of the plurality of range and angle combinations, the encoded vector includes the reflection intensities of a predetermined number of maximum reflection intensity peaks, the frequencies of the reflection intensities, and the widths of the reflection intensity peaks;

[0020] For each of the plurality of range and angle combinations, generating an encoded vector data set from the encoded vector;

[0021] Performing feature extraction on the encoded vector data set using a feature extraction neural network to generate an extracted feature data set with the extracted feature information for each of the plurality of range and angle combinations; and

[0022] Performing object detection on the extracted feature data set using an object detection neural network to generate an object detection output for identifying an object from the radar tensor.

[0023] Solution 2. The method according to Solution 1, wherein the predetermined number of reflection intensity peaks includes at least three reflection intensity peaks and no more than five reflection intensity peaks.

[0024] Solution 3. The method according to Solution 1, wherein the encoded vector is sorted in descending order of reflection intensity.

[0025] Solution 4. The method according to Solution 1, wherein a predetermined number of maximum reflection intensity peaks are based on the angles of the radar system used to capture the radar tensor.

[0026] Solution 5. The method according to Solution 1, including generating a frequency data set for each of the plurality of range and angle combinations, the frequency data set having the frequencies of the reflection intensities of each of the predetermined number of maximum reflection intensity peaks.

[0027] Solution 6. The method according to Solution 5, wherein the frequency data set is concatenated with the extracted feature data set to generate a concatenated data set.

[0028] Solution 7. The method according to Solution 6, wherein performing object detection on the extracted feature data set includes performing object detection on the concatenated data set including both the extracted feature data set and the frequency data set.

[0029] Solution 8. The method according to Solution 1, including determining the accuracy of the object detection neural network by comparing the object detection output with the ground truth data set of the radar tensor using a loss function.

[0030] Solution 9. The method according to Solution 8, including adapting the parameters of the object detection neural network based on the loss function to generate an updated object detection neural network.

[0031] Solution 10. The method according to Solution 8, including adapting the parameters of the feature extraction neural network based on the loss function to generate an updated object feature extraction neural network.

[0032] Solution 11. A non-transitory computer-readable storage medium containing programming instructions that are operable, when executed by a processor, to perform a method including:

[0033] Receiving a radar tensor, wherein the radar tensor includes the Doppler of a plurality of range and angle combinations;

[0034] Generating an encoding vector for each of the plurality of range and angle combinations from the radar tensor, wherein for each of the plurality of range and angle combinations, the encoding vector includes the reflection intensities of a predetermined number of maximum reflection intensity peaks, the frequencies of the reflection intensities, and the widths of the reflection intensity peaks;

[0035] Generating an encoding vector data set from the encoding vectors for each of the plurality of range and angle combinations;

[0036] Perform feature extraction on the encoded vector dataset using a feature extraction neural network to generate an extracted feature dataset with extracted feature information for each of a plurality of range and angle combinations; and

[0037] Perform object detection on the extracted feature dataset using an object detection neural network to generate an object detection output for identifying objects from the radar tensor.

[0038] Scheme 12. The computer-readable storage medium according to Scheme 11, wherein the predetermined number of reflection intensity peaks includes at least three reflection intensity peaks and no more than five reflection intensity peaks.

[0039] Scheme 13. The computer-readable storage medium according to Scheme 11, including generating a frequency dataset for each of the plurality of range and angle combinations, the frequency dataset having the frequency of the reflection intensity of each of the predetermined number of maximum reflection intensity peaks.

[0040] Scheme 14. The computer-readable storage medium according to Scheme 13, wherein the frequency dataset is connected to the extracted feature dataset to generate a connected dataset.

[0041] Scheme 15. The computer-readable storage medium according to Scheme 14, wherein performing object detection on the extracted feature dataset includes performing object detection on the connected dataset including both the extracted feature dataset and the frequency dataset.

[0042] Scheme 16. The computer-readable storage medium according to Scheme 11, including determining the accuracy of the object detection neural network by comparing the object detection output with the ground truth dataset of the radar tensor using a loss function.

[0043] Scheme 17. The computer-readable storage medium according to Scheme 16, including adapting the parameters of the object detection neural network based on the loss function to generate an updated object detection neural network.

[0044] Scheme 18. The computer-readable storage medium according to Scheme 16, including adapting the parameters of the feature extraction neural network based on the loss function to generate an updated object feature extraction neural network.

[0045] Scheme 19. A vehicle system, comprising:

[0046] At least one radar sensor configured to capture information for generating a radar tensor;

[0047] A controller in communication with the at least one radar sensor, wherein the controller is configured to:

[0048] Receive a radar tensor, where the radar tensor includes Doppler for a plurality of range and angle combinations;

[0049] Generate an encoded vector for each of the plurality of range and angle combinations from the radar tensor, where for each of the plurality of range and angle combinations, the encoded vector includes the reflection intensities of a predetermined number of maximum reflection intensity peaks, the frequencies of the reflection intensities, and the widths of the reflection intensity peaks;

[0050] Generate an encoded vector data set from the encoded vectors for each of the plurality of range and angle combinations;

[0051] Perform feature extraction on the encoded vector data set using a feature extraction neural network to generate an extracted feature data set with extracted feature information for each of the plurality of range and angle combinations; and

[0052] Perform object detection on the extracted feature data set using an object detection neural network to generate an object detection output for identifying an object from the radar tensor.

[0053] Scheme 20. The vehicle system according to Scheme 19, where the predetermined number of reflection intensity peaks includes at least three reflection intensity peaks and no more than five reflection intensity peaks. Brief Description of the Drawings

[0054] Figure 1 Illustrates an example vehicle and computer system according to the present disclosure.

[0055] Figure 2 Illustrates a flowchart of an example method for Doppler spectrum encoding for radar object detection.

[0056] Figure 3 Illustrates an example Doppler for a given range and distance.

[0057] Some embodiments of the present disclosure are now described only by way of example and with reference to the drawings. In all the drawings, the same reference numerals denote the same elements or elements of the same type. Detailed Description of the Embodiments

[0058] The present disclosure admits embodiments in many different forms. Representative examples of the present disclosure are shown in the drawings and described in detail herein as non-limiting examples of the disclosed principles. For this reason, elements and limitations described in the abstract, introduction, summary of the invention, and detailed description sections but not explicitly set forth in the claims should not be incorporated into the claims by implication, inference, or otherwise, either individually or jointly.

[0059] For the purposes of this specification, unless specifically disclaimed, the use of the singular includes the plural and vice versa, the terms "and" and "or" shall be both conjunctive and disjunctive, and the words "comprising", "including", "containing", "having", etc. shall mean "including without limitation". Additionally, approximate words such as "about", "almost", "substantially", "generally", "approximately", etc. may be used herein in the sense of "being at, close to, or nearly at" or "within 0-5% of" or "within acceptable manufacturing tolerances" or a logical combination thereof. As used herein, a component "configured to" perform a specified function can perform the specified function without modification, rather than merely having the potential to perform the specified function after further modification. In other words, when specifically configured to perform a specified function, the described hardware is specifically selected, created, implemented, utilized, programmed, and / or designed for the purpose of performing the specified function.

[0060] According to an exemplary embodiment, Figure 1 A vehicle 20 capable of operating in an autonomous mode or an automatic mode is shown. The vehicle 20 can be a fully autonomous vehicle or a semi-autonomous vehicle. The vehicle 20 includes a drive system 22 that controls the autonomous operation of the vehicle 20. The drive system 22 includes a sensor system 24 for obtaining information about the surroundings or environment of the vehicle 20; and a controller 26 for calculating possible actions of the autonomous vehicle based on the obtained information and for implementing one or more of the possible actions; and a human-machine interface 28 for communicating with an occupant (e.g., a driver or a passenger) of the vehicle 20. The sensor system 24 can include at least one depth sensor, such as a radar sensor 30.

[0061] The controller 26 can include processing circuitry that can include an application-specific integrated circuit (ASIC), an electronic circuit, a processor (shared, dedicated, or grouped) executing one or more software or firmware programs, and a memory, combinational logic circuitry, and / or other suitable components that provide the described functionality. The controller 26 can include a non-transitory computer-readable medium storing instructions that, when processed by one or more processors of the controller 26, implement a method 100 for encoding a radar Doppler spectrum for use in object detection described below.

[0062] In one example, the object detection of the method 100 is performed by the controller 26 on the vehicle 20 during an autonomous driving mode. In another example, a remote computer system 50 is used to train at least one of a feature extraction neural network or an object detection neural network, which will be described in more detail below. Although for simplicity of illustration, Figure 1The computer system 50 is depicted as a unitary computer module, but the computer system 50 can be physically implemented as one or more processing nodes having a non-transitory computer-readable storage medium 54, namely application-sufficient memory, and associated hardware and software, such as but not limited to a high-speed clock, timer, input / output circuitry, buffer circuitry, etc. The computer-readable storage medium 54 can include sufficient read-only memory, such as magnetic or optical memory. The computer-readable code or instructions for implementing the methods described below can be executed during the operation of the computer system 50. To this end, the computer system 50 can include one or more processors 52, such as logic circuits, application-specific integrated circuits (ASICs), central processing units, microprocessors, and / or other necessary hardware required to provide the programming functions described herein.

[0063] Figure 2 A flowchart of a method 100 for performing Doppler spectral encoding for radar object detection is illustrated. In one example, the method 100 performs object detection using a radar tensor 102 generated from information collected by a radar sensor 30 on a vehicle 20. In the illustrated example, the radar tensor 102 graphically represents the range 104 of the radar sensor 30 along a first dimension, the angle 106 of the radar sensor 30 along a second dimension, and the Doppler 108 along a third dimension. Although this disclosure describes the use of Doppler spectral encoding in automotive applications, this disclosure has applications in other fields, such as in industrial or security applications.

[0064] In the illustrated example, the method 100 begins by receiving the radar tensor 102. The radar tensor 102 can be received in real time from the radar sensor 30 on the vehicle 20 or be part of a collection of radar tensors 102 for training purposes. The radar tensor 102 represents a collection of Dopplers for a plurality of predetermined ranges 104 and angles 106. The Doppler 108 includes reflection intensity values taken at a plurality of different frequencies. In one example, the reflection intensity can be measured at up to 512 different frequencies.

[0065] In the method 100, the radar tensor 102 is received at block 110 to perform Doppler spectral encoding. One characteristic of Doppler spectral encoding is to reduce the size of the information captured in the radar tensor 102. This reduction in size allows for the object detection described herein to be performed with less processing power on the controller 26 or the computer system 50.

[0066] In one example, Doppler spectrum encoding begins by determining a predetermined number of reflection intensity peaks with relatively large reflection intensities for each of a plurality of combinations of range 104 and angle 106 in radar tensor 102. The number of reflection intensity peaks to be identified is based on the expected maximum number of objects that can be expected to be identified for a given range 104 and angle 106. For example, in the case of a smaller range, method 100 would be expected to identify fewer objects, while in the case of a larger angle, method 100 would be expected to identify a larger number of possible objects. Thus, in one example, the number of reflection intensity peaks to be identified can vary from greater than or equal to three and less than or equal to five. However, depending on the resolution of radar sensor 30, fewer than three or more than five reflection intensity peaks can be identified.

[0067] Using the predetermined number of reflection intensity peaks for a given range 104 and angle 106, method 100 will identify the corresponding number of maximum intensity peaks for each set of range 104 and angle 106. Figure 3 is a graphical representation 200 of Doppler 108 for a given range 104 and angle 106. Graphical representation 200 identifies frequency 204 along the x-axis and reflection intensity 206 along the y-axis. Line 208 represents the variation of reflection intensity as the frequency varies.

[0068] In addition, Figure 3 The illustrated example represents an example in the case where the predetermined number of reflection intensity peaks is equal to 3. For each of the identified maximum reflection intensity peaks, method 100 records the reflection intensities I1, I2, and I3, the frequencies f1, f2, and f3 associated with each of the reflection intensities I1, I2, and I3, and the widths Δ1, Δ2, and Δ3 of the reflection intensity peaks corresponding to the three maximum intensities I1, I2, and I3. The values identified above from Doppler spectrum encoding are stored in vector c, as shown in Equation 1 below.

[0069] c = {(I2, f2, Δ2), (I3, f3, Δ3), (I1, f1, Δ1)} Equation 1

[0070] As shown in the example Equation 1 above, the corresponding three sets of values are sorted in descending order with respect to the reflection intensity values Δ1, Δ2, and Δ3. However, the corresponding three sets of values can also be sorted in ascending order with respect to the reflection intensity values Δ1, Δ2, and Δ3.

[0071] Method 100 then generates an encoded vector data set 112 that represents an encoded vector c for each combination of range 104 and angle 106. The encoded vector data set 112 includes the range 104 along a first dimension, the angle 106 along a second dimension, and the encoded vector c 118 along a third dimension. One feature of the encoded vector data set 112 is that it is a data set with a greatly reduced size. For example, while the range 104 and angle 106 dimensions remain the same, the encoded vector c includes 9 values, as opposed to potentially hundreds of values for each combination of range 104 and angle 106.

[0072] Using the encoded vector data set 112, method 100 performs feature extraction on the encoded vector data set 112 at block 120 using a feature extraction neural network. The feature extraction neural network outputs an extracted feature data set 122. The extracted feature data set 122 includes the range 104 along a first dimension, the angle 106 along a second dimension, and the extracted feature information 119 along a third dimension. In the illustrated example, the feature extraction may include identifying specific features from the encoded vector data set 112, such as identifying lines or other shapes.

[0073] At block 124, the extracted feature data set 122 may optionally be concatenated with a frequency data set 130. In the illustrated example, the frequency data set 130 is generated by method 100 from the encoded vector data set 112. The frequency data set 130 includes frequencies f1, f2, and f3 associated with each of the corresponding reflection intensities I1, I2, and I3. When the frequency data set 130 is combined with the extracted feature data set 122, a concatenated data set 128 is generated by combining the frequency data set 130 into the extracted feature data set 122. One feature of the concatenated data set 128 is that it combines the frequencies f1, f2, and f3 with the extracted features for each combination of range 104 and angle 106.

[0074] Using the concatenated data set 128, method 100 proceeds to block 140 to perform object detection using an object detection neural network. In one example, the object detection neural network at block 140 has the option of using only the information from the extracted feature data set 122 or using the information from the frequency data set 130 in combination with the extracted feature data set 122. One feature of using the entire concatenated data set 128 with the object detection neural network at block 140 is associating specific frequencies with the extracted features in the extracted feature data set 122. Then, the object detection neural network at block 140 can provide an output that identifies objects from the original radar tensor 102.

[0075] Object detection performed by method 100 via block 140 can occur on controller 26 or remote computer system 50. This allows controller 26 to perform object detection in real time. However, for the purpose of training at least one of the feature extraction neural network from block 120 or the object detection neural network from block 140, remote computer system 50 can be utilized instead of controller 26. One feature of using remote computer system 50 for training purposes is that it reduces the computational power required by controller 26.

[0076] For training purposes, the output of the object detection neural network is fed into a loss function at block 150. At block 150, method 100 compares the output of the object detection neural network with the ground truth data set of radar tensor 102 from block 152.

[0077] The comparison results in an adaptation of the parameters used in at least one of the feature extraction neural network or the object detection neural network. The adaptation of the parameters is used to generate an updated feature extraction neural network and an updated object detection neural network, which can be used in subsequent training runs of method 100 or for object detection using method 100. Method 100 can be run multiple times with the same radar tensor 102 to extract the maximum learning potential from the radar tensor to adapt the parameters.

[0078] In addition, when method 100 is used to train the feature extraction neural network and the object detection neural network, a large data set of radar tensor 102 with the associated ground truth data set of radar tensor 102 can be used for further parameter adaptation of at least one of the feature extraction neural network or the object detection neural network. In addition, each of these additional radar tensors 102 used for training purposes can be run multiple times to extract the maximum learning from each radar tensor 102 and the associated ground truth data set.

[0079] Method 100 can continue using additional radar tensors 102 and ground truth data sets for training purposes until the loss function determines that the accuracy of the output of the object detection neural network from block 140 is within a predetermined accuracy of the ground truth data set. At this point, the supervised training of the feature extraction neural network and the object detection neural network is complete. Method 100 can run on controller 26 of vehicle 20 to perform object detection in real time to assist in autonomously navigating vehicle 20.

[0080] The terms "a" and "an" do not denote a limitation of quantity, but rather denote the presence of at least one of the referenced items. The term "or" means "and / or" unless clearly indicated otherwise by the context. References throughout the specification to "one aspect" mean that a particular element (e.g., a feature, a structure, a step, or a property) described in connection with that aspect is included in at least one aspect described herein, and may or may not be present in other aspects. In addition, it is to be understood that the described elements may be combined in a suitable manner in different aspects.

[0081] When an element such as a layer, a film, a region, or a substrate is referred to as being "on" another element, it can be directly on the other element or intervening elements may also be present. In contrast, when an element is referred to as being "directly on" another element, no intervening elements are present.

[0082] Unless otherwise specified herein, the test standards are the latest standards effective as of the filing date of the present application, or if priority is claimed, the filing date of the earliest priority application in which the test standards appear.

[0083] Unless otherwise defined, the technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure pertains.

[0084] Although the foregoing disclosure has been described in terms of exemplary embodiments, those skilled in the art will understand that various changes can be made without departing from its scope, and equivalents can be substituted for its elements. In addition, many modifications can be made to adapt a particular situation or material to the teachings of the disclosure without departing from its scope. Accordingly, it is intended that the disclosure not be limited to the particular embodiments disclosed, but will include embodiments falling within its scope.

Claims

1. A method for performing Doppler spectrum encoding for radar object detection, the method comprising: Receiving a radar tensor, wherein the radar tensor includes Doppler for a plurality of range and angle combinations; Generating, for each of the plurality of range and angle combinations, an encoding vector from the radar tensor, wherein for each of the plurality of range and angle combinations, the encoding vector includes the reflection intensity of a predetermined number of maximum reflection intensity peaks, the frequency of the reflection intensity, and the width of the reflection intensity peaks; Generating, for each of the plurality of range and angle combinations, an encoding vector data set from the encoding vector; Performing feature extraction on the encoding vector data set with a feature extraction neural network to generate an extracted feature data set having extracted feature information for each of the plurality of range and angle combinations; And Performing object detection on the extracted feature data set with an object detection neural network to generate an object detection output for identifying an object from the radar tensor.

2. The method according to claim 1, wherein the predetermined number of reflection intensity peaks includes at least three reflection intensity peaks and not more than five reflection intensity peaks.

3. The method according to claim 1, wherein the encoding vectors are sorted in descending order of reflection intensity.

4. The method according to claim 1, wherein the predetermined number of maximum reflection intensity peaks is based on the angle of the radar system used to capture the radar tensor.

5. The method according to claim 1, comprising generating, for each of the plurality of range and angle combinations, a frequency data set having the frequency of the reflection intensity of each of the predetermined number of maximum reflection intensity peaks.

6. The method according to claim 5, wherein the frequency data set is concatenated to the extracted feature data set to generate a concatenated data set.

7. The method according to claim 6, wherein performing object detection on the extracted feature data set includes performing object detection on the concatenated data set including both the extracted feature data set and the frequency data set.

8. The method according to claim 1, comprising determining the accuracy of the object detection neural network by comparing the object detection output with a ground truth data set of the radar tensor using a loss function.

9. The method according to claim 8, comprising adapting the parameters of the object detection neural network based on the loss function to generate an updated object detection neural network.

10. The method according to claim 8, comprising adapting the parameters of the feature extraction neural network based on the loss function to generate an updated object feature extraction neural network.

Images