Target person tracking method, computer equipment and readable storage medium

Abstract

The invention provides a target person tracking method, computer equipment and a readable storage medium, and relates to the technical field of computer vision. The method comprises the following steps: calling a human body part feature detection model to extract a first local feature vector of a preset human body part actually existing in a real-time figure image of a to-be-detected figure and a second local feature vector of a preset human body part actually existing in a reference figure image of a target figure; at least one target human body part exists in the real-time figure image and the reference figure image; and when the feature similarity between a first target feature vector obtained by fusing the first local feature vectors of all the target human body parts and a second target feature vector obtained by fusing the second local feature vectors of all the target human body parts exceeds a preset similarity threshold, judging that the to-be-detected person is the target person. Therefore, the effect of low computing resource loss is ensured, the problem of misalignment of matching features is effectively solved, and the target person tracking accuracy is improved.

Description

Technical Field

[0001] The present application relates to the field of computer vision technology, and in particular to a target person tracking method, a computer device, and a readable storage medium. Background Art

[0002] With the continuous development of science and technology, computer vision technology is increasingly used in video surveillance, social security and other fields. Robot security is an important research direction of computer vision technology. In order to ensure that robots can assist humans in completing security protection work, robots are usually required to have the ability to recognize and follow sensitive people in security sites in real time, so that robots can send warning signals of sensitive people to the control center in time. Therefore, in the process of robots following the target person, the accuracy of the robot's recognition of the target person is an important factor affecting the accuracy of the robot's following.

[0003] In the process of the robot following the target person, there is often a situation where the target person walks out of the robot's monitoring field of view and then walks back into the monitoring field of view. It is necessary to extract features of all people in the monitoring field of view and match them with the saved human features of the target person, so that the target person can be accurately found among many people when he walks back into the monitoring field of view.

[0004] However, it is worth noting that during the robot following process, due to the limitations of the robot's own camera installation height and camera viewing angle, as the target person is closer or farther away from the robot, the target person's body parts that appear in the real-time image taken by the robot's camera will also be different (for example, in the real-time image taken by the robot, the target person may only appear in any one or more combinations of human body parts such as the head, arms, thighs, torso, upper body, lower body, left body and right body, or the target person may appear directly with the whole body). At this time, if the global features in the real-time image are directly matched with the saved human body features of the target person, there will obviously be a problem of feature misalignment (for example, the global features of the half body and the global features of the whole body are obviously not aligned), resulting in the final person recognition result being extremely prone to misjudgment, and unable to achieve high-precision tracking of the target person. Summary of the invention

[0005] In view of this, the purpose of the present application is to provide a target person tracking method, a computer device and a readable storage medium, which can extract local features of human body parts of matching types for the image of the person to be detected and the image of the target person for feature matching on the basis of ensuring low computing resource loss, so as to effectively solve the problem of misalignment of matching features, improve the accuracy of identifying the target person, and facilitate improving the accuracy of the robot's person following.

[0006] In order to achieve the above purpose, the technical solution adopted in the embodiment of the present application is as follows:

[0007] In a first aspect, the present application provides a target person tracking method, the method comprising:

[0008] Acquire a real-time image of the person to be detected and a reference image of the target person;

[0009] Calling a pre-trained human body part feature detection model to extract a first local feature vector of at least one preset human body part actually existing in the real-time human image, and a second local feature vector of at least one preset human body part actually existing in the reference human image;

[0010] Detecting whether the real-time character image and the reference character image have at least one target human body part in common based on the extracted first local feature vector and the second local feature vector;

[0011] If it is detected that the real-time person image and the reference person image do not have at least one target human body part in common, it is determined that the person to be detected is not the target person, otherwise it is detected whether the feature similarity between the first target feature vector and the second target feature vector exceeds a preset similarity threshold, wherein the first target feature vector is obtained by merging the first local feature vectors of all target human body parts, and the second target feature vector is obtained by merging the second local feature vectors of all target human body parts;

[0012] When it is detected that the corresponding feature similarity exceeds a preset similarity threshold, it is determined that the person to be detected is the target person, otherwise it is determined that the person to be detected is not the target person.

[0013] In an optional embodiment, the human body part feature detection model includes a backbone network and a plurality of main branch networks, wherein each main branch network is solely responsible for a local feature extraction operation of a preset human body part, and the step of calling the pre-trained human body part feature detection model to extract a first local feature vector of at least one preset human body part actually existing in the real-time human image, and a second local feature vector of at least one preset human body part actually existing in the reference human image, comprises:

[0014] For each person image in the real-time person image and the reference person image, calling the backbone network to extract image features of the person image to obtain a global feature map of the person image;

[0015] The global feature map of the person image is input to all main branch networks at the same time, and each main branch network is called to extract the actual local feature vector matching the responsible preset human body part from the input global feature map;

[0016] For each actual local feature vector of a preset human body part corresponding to the person image, detecting whether the actual local feature vector of the preset human body part is valid;

[0017] When it is detected that the actual local feature vector of the preset human body part is valid, the actual local feature vector of the preset human body part is used as a target local feature vector actually existing in the character image, wherein the target local feature vector is a first local feature vector actually existing in the real-time character image, or the target local feature vector is a second local feature vector actually existing in the reference character image.

[0018] In an optional implementation, the step of detecting whether the actual local feature vector of each preset human body part corresponding to the character image is valid comprises:

[0019] Calculating the vector modulus of the actual local feature vector of the preset human body part;

[0020] Detecting whether the calculated vector modulus exceeds a preset modulus threshold value that matches the preset human body part, wherein different preset human body parts correspond to different preset modulus threshold values;

[0021] If it is detected that the vector modulus exceeds a preset modulus threshold matching the preset human body part, the actual local feature vector of the preset human body part is determined to be valid; otherwise, the actual local feature vector of the preset human body part is determined to be invalid.

[0022] In an optional embodiment, the step of detecting whether the real-time person image and the reference person image have at least one target human body part in common based on the extracted first local feature vector and the second local feature vector comprises:

[0023] Performing intersection processing on the preset human body parts corresponding to all the first local feature vectors and the preset human body parts corresponding to all the second local feature vectors to obtain a corresponding human body part result set;

[0024] When the human body part result set is an empty set, determining that the real-time human image and the reference human image do not have at least one target human body part in common;

[0025] When the human body part result set is a non-empty set, it is determined that the real-time human image and the reference human image have at least one target human body part in common, and all preset human body parts in the human body part result set are respectively regarded as a target human body part.

[0026] In an optional embodiment, the method further comprises:

[0027] Acquire a human image training sample set for a human body detection model to be trained, and human body part segmentation results of the human image training sample set that match multiple preset human body parts;

[0028] Calling the backbone network in the human body detection model to be trained to extract image features from the human image training sample set to obtain a corresponding global feature atlas;

[0029] Calling each main branch network in the human body detection model to be trained to extract local features from the global feature atlas, respectively, to obtain a first local feature output vector set of each main branch network;

[0030] Calling each auxiliary branch network in the human body detection model to be trained to extract a local feature atlas matching the responsible preset human body part from the global feature atlas according to the human body part segmentation result of the human image training sample set, and performing local feature extraction on the local feature atlas to obtain a second local feature output vector set of each auxiliary branch network, wherein each auxiliary branch network in the human body detection model to be trained is solely responsible for the local feature extraction operation of one preset human body part among the multiple preset human body parts, and each auxiliary branch network corresponds to a main branch network and shares convolution layer weight parameters with the main branch network;

[0031] For each main branch network, the main branch network is optimized according to the second local feature output vector set and the first local feature output vector set corresponding to the main branch network, and then the step of calling the backbone network to extract image features from the person image training sample set is continued until the model performance of the to-be-trained human body detection model reaches the expected performance state;

[0032] All auxiliary branch networks in the trained human body detection model to be trained are pruned to obtain the human body part feature detection model.

[0033] In an optional implementation, the step of obtaining the human body part segmentation results matching multiple preset human body parts of the human image training sample set includes:

[0034] For each sample person image in the person image training sample set, calling a human body part segmentation model to perform part distribution recognition in the sample person image according to the multiple preset human body parts, and obtaining actual distribution data of each of the multiple preset human body parts in the sample person image;

[0035] The actual distribution data corresponding to the plurality of preset human body parts of all sample human images in the human image training sample set are sorted to obtain the human body part segmentation result of the human image training sample set.

[0036] In an optional embodiment, for each main branch network, the step of performing network optimization on the main branch network according to the second local feature output vector set and the first local feature output vector set corresponding to the main branch network includes:

[0037] For each sample person image in the person image training sample set, determine a second local feature output vector and a first local feature output vector that match the sample person image from the second local feature output vector set and the first local feature output vector set corresponding to the main branch network;

[0038] Calculating a feature output vector difference between a second local feature output vector and a first local feature output vector that match the sample person image;

[0039] With the purpose of minimizing the difference of the feature output vector, the convolutional layer weight parameters of the main branch network are optimized.

[0040] In an optional implementation, the difference between the feature output vectors of the kth main branch network and the kth auxiliary branch network in the human detection model to be trained for the same sample person image is calculated using the following formula:

[0041]

[0042] Among them, C k It is used to represent the difference between the feature output vectors of the kth main branch network and the kth auxiliary branch network for the same sample person image, h k The first local feature output vector of the kth main branch network for the sample person image, g k The second local feature output vector of the kth auxiliary branch network for the sample person image.

[0043] In a second aspect, the present application provides a computer device, including a processor and a memory, wherein the memory stores a computer program that can be executed by the processor, and the processor can execute the computer program to implement the target person tracking method described in any one of the aforementioned embodiments.

[0044] In a third aspect, the present application provides a readable storage medium having a computer program stored thereon, and when the computer program is executed by a computer device, the target person tracking method described in any one of the aforementioned embodiments is implemented.

[0045] In this case, the beneficial effects of the embodiments of the present application may include the following:

[0046] The present application calls a human body part feature detection model to extract a first local feature vector of at least one preset human body part actually existing in the real-time human image of the person to be detected, and a second local feature vector of at least one preset human body part actually existing in the reference human image of the target person. The person to be detected is determined to be the target person only when at least one target human body part coexists in the real-time human image and the reference human image, and the feature similarity between the first target feature vector obtained by fusion of the first local feature vectors of all target human body parts and the second target feature vector obtained by fusion of the second local feature vectors of all target human body parts exceeds a preset similarity threshold. Otherwise, the person to be detected is directly determined not to be the target person. In this way, on the basis of utilizing a single network model to realize the local feature extraction function of different human body parts and ensuring the low computing resource loss effect of the person tracking operation, the local features of human body parts of matching types are extracted for the image of the person to be detected and the image of the target person for feature matching, so as to solve the problem of misalignment of matching features in the person tracking process, improve the recognition accuracy of the target person, and facilitate improving the person following accuracy of the robot.

[0047] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, preferred embodiments are specifically cited below and described in detail with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0049] Figure 1 This is a schematic diagram of the PCB model architecture of the pedestrian local feature re-identification model;

[0050] Figure 2 A schematic diagram of the composition of a computer device provided in an embodiment of the present application;

[0051] Figure 3 One of the flow charts of the target person tracking method provided in the embodiment of the present application;

[0052] Figure 4 for Figure 3 A schematic flow chart of the sub-steps included in step S220;

[0053] Figure 5The second flowchart of the target person tracking method provided in the embodiment of the present application;

[0054] Figure 6 This is a schematic diagram of the training architecture of the human body detection model to be trained according to an embodiment of the present application.

[0055] Icons: 10 - computer device; 11 - memory; 12 - processor; 13 - communication unit. DETAILED DESCRIPTION

[0056] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings here can be arranged and designed in various different configurations.

[0057] Therefore, the following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the present application for which protection is sought, but merely represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present application.

[0058] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0059] In the description of the present application, it should be understood that the terms "center", "up", "down", "left", "right", "vertical", "horizontal", "inside", "outside", etc. indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, or are the orientations or positional relationships in which the product of the application is conventionally placed when in use, or are the orientations or positional relationships conventionally understood by those skilled in the art. They are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present application.

[0060] In the description of this application, it should also be noted that, unless otherwise clearly specified and limited, the terms "set", "install", "connect", and "connect" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two elements. For ordinary technicians in this field, the specific meanings of the above terms in this application can be understood according to specific circumstances.

[0061] In addition, in the description of the present application, it is understood that the relational terms such as the term "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that there is any such actual relationship or order between these entities or operations. Moreover, the term "include", "comprise" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements includes not only those elements, but also includes other elements that are not clearly listed, or also includes elements inherent to such process, method, article or equipment. In the absence of more restrictions, the elements defined by the sentence "comprise one..." do not exclude the existence of other identical elements in the process, method, article or equipment including the elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present application can be understood by specific circumstances.

[0062] The applicant has found through hard research that the problem of feature misalignment during character image matching can be solved by extracting and matching local features. The local feature extraction and matching solutions currently used by the industry for person re-identification operations mainly include the following two:

[0063] Option 1: Use Figure 1The PCB (Part-based Convolutional Baseline) model architecture shown directly constructs a pedestrian local feature re-identification model assuming that the pedestrian is standing. The PCB model architecture directly divides the global feature map of the human body into K sub-parts from top to bottom. Taking K=2 as an example, each sub-part extracts a local feature vector with a vector dimension of 1*N. The global feature is a feature vector composed of the local feature vectors of the K sub-parts (whose vector dimension is 1*N). During training, it is necessary to ensure that the global features and the local features of the K sub-parts are similar. Since the sub-part division method of the model architecture is relatively simple, the following problems often occur: (1) During model training, the training set may contain incomplete human body pictures. Therefore, the training process may force the upper half of the incomplete human body picture to be similar to the upper half of the complete human body picture. This is obviously an unreasonable constraint, which will lead to poor performance of the final trained model and the problem of misaligned training features. (2) When using the final trained pedestrian local feature re-identification model for matching, the upper body picture is input into the model. At this time, the output global feature is the upper body feature vector (vector dimension is 1*2N), and the full body picture is input into the model. The corresponding output global feature is the full body feature vector (vector dimension is 1*2N). If the feature vectors of the upper body in the two pictures need to be matched, the feature vectors with a vector dimension of 1*2N and the feature vectors with a vector dimension of 1*N need to be matched. However, this matching method cannot calculate the similarity and there is a problem of mismatched feature dimensions. If the global features of the two pictures are directly matched, there will be a problem of misaligned matching features. In other words, the PCB model architecture is actually more suitable for feature extraction and matching of full-body images of pedestrians. It cannot effectively solve the problem of misaligned matching features, and cannot perform high-precision tracking of the target person as the target person is at different distances from the robot.

[0064] Solution 2: Train a local feature extractor for different parts of the human body, and then use multiple local feature extractors to collaboratively extract local features of different parts of the same human body image, and then combine the local features of multiple parts to obtain the overall features of the corresponding human body image, so as to solve the problem of misaligned matching features in the target person tracking process by matching the overall features of different human body images. However, it is worth noting that this solution requires a lot of computing power to detect different parts of the human body and involves complex post-processing processes. In the process of the end-side robot following, low resource consumption and high-speed computing are factors that need to be considered. Therefore, this solution is not actually suitable for robot following operations.

[0065] In this case, in order to solve the above problems, the embodiments of the present application provide a target person tracking method, a computer device and a readable storage medium, which can extract local features of human body parts of matching types for the image of the person to be detected and the target person image respectively for feature matching on the basis of ensuring low computing resource loss, so as to effectively solve the problem of misalignment of matching features and mismatch of feature dimensions, improve the accuracy of identifying the target person, and facilitate improving the accuracy and timeliness of the robot's person following.

[0066] In conjunction with the accompanying drawings, some embodiments of the present application are described in detail below. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.

[0067] Please refer to Figure 2 , Figure 2 1 is a schematic diagram of the composition of the computer device 10 provided in the embodiment of the present application. In the embodiment of the present application, the computer device 10 can be connected to the mobile robot for communication, and obtain any character image of the target character that the mobile robot expects to follow, as well as any one or more real-time character images of the characters to be detected captured by the mobile robot in real time through the camera, and then based on the human body part feature detection model carried by itself, the local features of the human body parts of the characters to be detected (i.e., the real-time character image of the characters to be detected) and the target character image (i.e., any character image of the target character) are extracted for feature matching, thereby using a single neural network model to realize the local feature extraction function of different human body parts, ensuring the low computing resource loss effect of the character tracking operation, effectively solving the problem of misalignment of matching features and mismatch of feature dimensions in the character tracking process, improving the recognition accuracy of the target character, and facilitating the improvement of the accuracy and timeliness of the character following of the mobile robot. Among them, the target person can be a specific person selected by the user, or a person recognized by the mobile robot itself that meets specific requirements (for example, specific facial feature rules, abnormal behavior detection rules, specific posture detection rules); the mobile robot can be, but is not limited to, a humanoid robot, a tracked robot, an aerial photography robot, etc.; the computer device 10 can be a computing device independent of the mobile robot, and the computing device can be, but is not limited to, a tablet computer, a laptop computer, a server, a personal computer, etc.; the computer device 10 can also be integrated with the mobile robot.

[0068] In the embodiment of the present application, the computer device 10 may include a memory 11, a processor 12, and a communication unit 13. The memory 11, the processor 12, and the communication unit 13 are electrically connected to each other directly or indirectly to achieve data transmission or interaction. For example, the memory 11, the processor 12, and the communication unit 13 may be electrically connected to each other via one or more communication buses or signal lines.

[0069] In this embodiment, the memory 11 may be, but not limited to, a random access memory (RAM), a read only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable read-only memory (EEPROM), etc. The memory 11 is used to store a computer program, and the processor 12 may execute the computer program accordingly after receiving an execution instruction.

[0070] In this embodiment, the processor 12 may be an integrated circuit chip with signal processing capability. The processor 12 may be a general-purpose processor, including a central processing unit (CPU), a graphics processing unit (GPU), a network processor (NP), a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), or at least one of other programmable logic devices, discrete gate or transistor logic devices, and discrete hardware components. A general-purpose processor may be a microprocessor or the processor may be any conventional processor, etc., which may implement or execute the disclosed methods, steps, and logic block diagrams in the embodiments of the present application.

[0071] In this embodiment, the communication unit 13 is used to establish a communication connection between the computer device 10 and other electronic devices through a network, and to send and receive data through the network, wherein the network includes a wired communication network and a wireless communication network. For example, the computer device 10 can obtain the real-time image of any one or more persons to be detected captured by the mobile robot in real time, and any image of the target person that the mobile robot wants to follow, through the communication unit 13, and then feed back the image feature matching result of the person to be detected and the target person to the mobile robot, so as to improve the accuracy and timeliness of the mobile robot's following of the target person.

[0072] In an embodiment of the present application, the computer device 10 may pre-store a specific computer program related to the target person tracking function in the memory 11, and by driving the processor 12 to execute the specific computer program stored in the memory 11, on the basis of ensuring low computing resource loss, extract local features of human body parts of matching types for the image of the person to be detected and the image of the target person for feature matching, so as to effectively solve the problem of misalignment of matching features and mismatch of feature dimensions, improve the accuracy of identifying the target person, and facilitate improving the accuracy and timeliness of the mobile robot's person following.

[0073] Understandably, Figure 2 The block diagram shown is only a schematic diagram of a composition of the computer device 10. The computer device 10 may also include Figure 2 More or fewer components as shown, or with Figure 2 Different configurations are shown. Figure 2 Each component shown in the figure can be implemented by hardware, software or a combination thereof.

[0074] In the present application, in order to ensure that the computer device 10 can solve the problem of misalignment of matching features and mismatch of feature dimensions in the process of person tracking on the basis of ensuring low computing resource consumption, improve the accuracy of identifying the target person, and facilitate improving the accuracy and timeliness of the robot's person following, the present application embodiment provides a method for tracking a target person to achieve the above purpose. The target person tracking method provided by the present application is described in detail below.

[0075] Please refer to Figure 3 , Figure 3 is one of the flow charts of the target person tracking method provided in the embodiment of the present application. In the embodiment of the present application, Figure 3 The target person tracking method shown may include steps S210 to S260.

[0076] Step S210, obtaining a real-time image of the person to be detected and a reference image of the target person.

[0077] In this embodiment, the real-time character image of the person to be detected can be collected in real time by the mobile robot and sent to the computer device 10; the reference character image of the target person can be the actual character image of the target person taken by the mobile robot during the character following process, or it can be the actual character image of the target person pre-stored in the computer device 10. Generally speaking, the reference character image of the target person should use the full-body character image of the target person as much as possible to ensure the accuracy of identifying the target person.

[0078] Step S220, calling a pre-trained human body part feature detection model to extract a first local feature vector of at least one preset human body part actually existing in the real-time human image, and a second local feature vector of at least one preset human body part actually existing in the reference human image.

[0079] In this embodiment, the human body part feature detection model includes a backbone network and multiple main branch networks, wherein each main branch network is solely responsible for a local feature extraction operation of a preset human body part, and the backbone network is used to perform global feature extraction on the input character image; the number of main branch networks present in the human body part feature detection model represents the number of preset human body parts that the human body part feature detection model can perform local feature extraction on; the multiple preset human body parts involved in the human body part feature detection model may include but are not limited to: head, arms, thighs, torso, upper body, lower body, left body and right body parts; the dimensions of the local feature vectors extracted by the main branch networks included in the human body part feature detection model remain consistent (for example, a feature vector dimension of 1*N).

[0080] The computer device 10 can input the real-time character image and the reference character image into the human body part feature detection model respectively, and the human body part feature detection model can extract local features of different preset human body parts for the input character image, thereby obtaining a local feature vector of at least one preset human body part actually existing in the corresponding character image, that is, obtaining a first local feature vector of at least one preset human body part actually existing in the real-time character image, and a second local feature vector of at least one preset human body part actually existing in the reference character image.

[0081] Optionally, see Figure 4 , Figure 4 yes Figure 3In the embodiment of the present application, the step S220 may include sub-steps S221 to S224 to effectively extract the real local feature vectors of all preset human body parts actually existing in any human image.

[0082] Sub-step S221, for each person image in the real-time person image and the reference person image, calling the backbone network to extract image features of the person image to obtain a global feature map of the person image.

[0083] In sub-step S222, the global feature map of the character image is input to all main branch networks at the same time, and each main branch network is called to extract the actual local feature vector that matches the responsible preset human body part from the input global feature map.

[0084] Sub-step S223, for each actual local feature vector of a preset human body part corresponding to the character image, detecting whether the actual local feature vector of the preset human body part is valid.

[0085] In this embodiment, for any person image, if a certain preset human body part does not actually exist in the person image, then each vector element in the actual local feature vector of the preset human body part in the person image is basically 0, so the vector modulus of the actual local feature vector can be calculated to assist in judging whether the corresponding actual local feature vector is valid, and then determine whether the preset human body part exists in the person image. Among them, if the actual local feature vector of a certain preset human body part in a certain person image is invalid, it can be determined that the preset human body part does not actually exist in the person image.

[0086] Specifically, the step of detecting whether the actual local feature vector of each preset human body part corresponding to the character image is valid may include:

[0087] Calculating the vector modulus of the actual local feature vector of the preset human body part;

[0088] Detecting whether the calculated vector modulus exceeds a preset modulus threshold value that matches the preset human body part, wherein different preset human body parts correspond to different preset modulus threshold values;

[0089] If it is detected that the vector modulus exceeds a preset modulus threshold matching the preset human body part, the actual local feature vector of the preset human body part is determined to be valid; otherwise, the actual local feature vector of the preset human body part is determined to be invalid.

[0090] Therefore, the present application can effectively confirm the preset human body parts composition actually existing in any character image by executing the specific step process of the above-mentioned sub-step S223.

[0091] Sub-step S224, when it is detected that the actual local feature vector of the preset human body part is valid, the actual local feature vector of the preset human body part is used as a target local feature vector actually existing in the person image.

[0092] Among them, for the real-time character image of the person to be detected, the corresponding target local feature vector is a first local feature vector of a preset human body part that actually exists in the real-time character image; for the reference character image of the target person, the corresponding target local feature vector is a second local feature vector of a preset human body part that actually exists in the reference character image.

[0093] Therefore, the present application can effectively extract the real local feature vectors of all preset human body parts actually existing in any person image from the global feature map of the person image by executing the above sub-steps S221 to S224.

[0094] Step S230: Detect whether the real-time person image and the reference person image have at least one target human body part in common based on the extracted first local feature vector and the second local feature vector.

[0095] In this embodiment, the target human body part is a preset human body part that actually exists in both the real-time human image and the reference human image. The computer device 10 can determine the preset human body part composition that actually exists in the real-time human image based on all the first local feature vectors extracted from the real-time human image, and at the same time determine the preset human body part composition that actually exists in the reference human image based on all the second local feature vectors extracted from the reference human image, and then determine whether the real-time human image and the reference human image have at least one target human body part in common by comparing the preset human body part compositions corresponding to the real-time human image and the reference human image.

[0096] Among them, when it is determined that the real-time person image and the reference person image do not have at least one target human body part in common, it indicates that the person image to be detected and the target person image do not have local features of human body parts of matching types, and the feature similarity between the person image to be detected and the target person image can be directly determined to be 0. At this time, the computer device 10 will execute step S260 accordingly; when it is determined that the real-time person image and the reference person image have at least one target human body part in common, it indicates that the person image to be detected and the target person image have local features of human body parts of matching types, and the real local feature vectors of the target human body parts that exist in both the person image to be detected and the target person image can be directly feature matched to solve the problem of misalignment of matching features and mismatch of feature dimensions in the person tracking process. At this time, the computer device 10 will execute step S240 accordingly.

[0097] Optionally, in the embodiment of the present application, step S230 may include:

[0098] Performing intersection processing on the preset human body parts corresponding to all the first local feature vectors and the preset human body parts corresponding to all the second local feature vectors to obtain a corresponding human body part result set;

[0099] When the human body part result set is an empty set, determining that the real-time human image and the reference human image do not have at least one target human body part in common;

[0100] When the human body part result set is a non-empty set, it is determined that the real-time human image and the reference human image have at least one target human body part in common, and all preset human body parts in the human body part result set are respectively regarded as a target human body part.

[0101] Therefore, the present application can effectively confirm whether there are local features of human body parts of matching types between the image of the person to be detected and the image of the target person by executing the specific step process of the above-mentioned step S230.

[0102] Step S240: Detect whether the feature similarity between the first target feature vector and the second target feature vector exceeds a preset similarity threshold.

[0103] In this embodiment, the first target feature vector is obtained by fusing the first local feature vectors of all target human body parts, and the second target feature vector is obtained by fusing the second local feature vectors of all target human body parts, wherein the local feature vector fusion method of the first target feature vector and the second target feature vector are consistent to ensure that the first target feature vector and the second target feature vector have the same feature vector dimension, so as to solve the matching feature misalignment problem and feature dimension mismatch problem related to the feature matching operation in the person tracking process.

[0104] The computer device 10 can calculate the feature similarity between the first target feature vector and the second target feature vector, and determine whether the calculated feature similarity exceeds a preset similarity threshold, so as to realize the feature matching operation of the local features of the human body parts that exist simultaneously in the image of the person to be detected and the image of the target person. When it is determined that the calculated feature similarity exceeds the preset similarity threshold, it indicates that the local features of the human body parts that exist simultaneously in the image of the person to be detected and the image of the target person are successfully matched, and it can be basically determined that the person to be detected and the target person belong to the same person, and at this time, the computer device 10 will correspondingly execute step S250; when it is determined that the calculated feature similarity does not exceed the preset similarity threshold, it indicates that the local features of the human body parts that exist simultaneously in the image of the person to be detected and the image of the target person fail to match, and it can be basically determined that the person to be detected and the target person do not belong to the same person, and at this time, the computer device 10 will correspondingly execute step S260.

[0105] Step S250, determining that the person to be detected is a target person.

[0106] In this embodiment, when the computer device 10 determines that the person to be detected is the target person, the computer device 10 can inform the mobile robot of the corresponding person recognition result, so that the mobile robot can adjust its own posture based on the human body position information of the person to be detected in the corresponding real-time person image, ensuring that the mobile robot can continue to follow the target person, thereby effectively improving the accuracy and timeliness of the mobile robot's person following.

[0107] Step S260, determining that the person to be detected is not the target person.

[0108] In this embodiment, when the computer device 10 determines that the person to be detected is not the target person, the computer device 10 can inform the mobile robot of the corresponding person recognition result, so that the mobile robot can change the camera shooting object in time (that is, change the person to be detected), so as to quickly determine whether there is a target person within the detectable range of the mobile robot, and the actual location information of the target person when it exists, thereby effectively improving the accuracy and timeliness of the mobile robot's person following.

[0109] Therefore, the present application can, by executing the above steps S210 to S260, utilize a single neural network model to realize the local feature extraction function of different human body parts, thereby ensuring the low computing resource consumption effect of the character tracking operation, effectively solve the matching feature misalignment problem and feature dimension mismatch problem in the character tracking process, improve the recognition accuracy of the target person, and facilitate improving the character following accuracy and character following timeliness of the mobile robot.

[0110] Optionally, see Figure 5 , Figure 5 This is the second flow chart of the target person tracking method provided in the embodiment of the present application. Figure 3 Compared with the target person tracking method shown in Figure 5 The target person tracking method shown may also include steps S310 to S370 to ensure that the trained human body part feature detection model can realize the local feature extraction function of different human body parts for any person image and effectively solve the problem of training feature misalignment during the model training process.

[0111] Step S310, obtaining a human image training sample set for a human detection model to be trained, and human body part segmentation results of the human image training sample set that match multiple preset human body parts.

[0112] In this embodiment, the character image training sample set includes multiple sample character images corresponding to multiple sample characters, and the multiple sample character images corresponding to the same sample character may include incomplete human body images or full-body images. The human body part segmentation result of the character image training sample set includes the actual distribution data of each of the multiple preset human body parts in each sample character image, and the actual distribution data may include information such as the image area, image area contour position, etc. of each preset human body part in the corresponding sample character image. The human body part segmentation result can be obtained by manual labeling, or by using a pre-trained human body part segmentation model to perform human body part segmentation processing on each sample character image in the character image training sample set.

[0113] Among them, it can be understood that when a certain preset human body part does not actually exist in a sample character image, the actual distribution data of the preset human body part in the corresponding sample character image can be set to a special state (for example, the image region area of ​​the preset human body part in the corresponding sample character image is set to 0, and the image region contour position of the preset human body part in the corresponding sample character image is recorded as empty) to indicate that the preset human body part does not exist in the corresponding sample character image.

[0114] In one implementation of this embodiment, when the human body part segmentation results of the human image training sample set are obtained by processing the human body part segmentation model, the step of "obtaining the human body part segmentation results of the human image training sample set that match multiple preset human body parts" in step S310 may include:

[0115] For each sample person image in the person image training sample set, calling a human body part segmentation model to perform part distribution recognition in the sample person image according to the multiple preset human body parts, and obtaining actual distribution data of each of the multiple preset human body parts in the sample person image;

[0116] The actual distribution data corresponding to the plurality of preset human body parts of all sample human images in the human image training sample set are sorted to obtain the human body part segmentation result of the human image training sample set.

[0117] Among them, it can be understood that when a certain preset human body part does not actually exist in a sample character image, the actual distribution data of the preset human body part in the corresponding sample character image can be set to a special state (for example, the image region area of ​​the preset human body part in the corresponding sample character image is set to 0, and the image region contour position of the preset human body part in the corresponding sample character image is recorded as empty) to indicate that the preset human body part does not exist in the corresponding sample character image.

[0118] Step S320, calling the backbone network in the human body detection model to be trained to extract image features from the human image training sample set to obtain a corresponding global feature atlas.

[0119] In this embodiment, please refer to Figure 6 The human body detection model to be trained provided in this application may include a backbone network, multiple main branch networks and multiple auxiliary branch networks ( Figure 6 Where n is greater than or equal to 2), wherein the backbone network is used to perform global feature extraction on the input character image, each main branch network is responsible for effectively extracting a local feature vector of a preset human body part from the global feature map extracted by the backbone network, and each auxiliary branch network is responsible for extracting a local feature vector of a preset human body part from the local feature map adapted by itself, and the local feature map adapted by each auxiliary branch network is a local feature map extracted from the aforementioned global feature map and matching the preset human body part that the auxiliary branch network is responsible for; wherein each auxiliary branch network corresponds to a main branch network separately and shares convolutional layer weight parameters with the main branch network, and each auxiliary branch network is used to assist the training of the corresponding main branch network, so that the corresponding main branch network can directly extract the local feature vector of the preset human body part that it is responsible for from the global feature map.

[0120] Therefore, after obtaining the person image training sample set, each sample person image in the person image training sample set can be input into the human body detection model to be trained, and the backbone network in the human body detection model to be trained can perform image feature extraction on the received sample person image to obtain the global feature map corresponding to the sample person image in the global feature map set.

[0121] Step S320, calling each main branch network in the human body detection model to be trained to perform local feature extraction on the global feature atlas respectively, to obtain the first local feature output vector set of each main branch network.

[0122] In this embodiment, after the skeleton network extracts the global feature map of any sample person image, the global feature map of the sample person image is synchronously input to each main branch network, so that each main branch network extracts a local feature output vector matching the preset human body part that it is responsible for from the received global feature map based on the convolution layer weight parameters shared with the corresponding auxiliary branch network, thereby obtaining the first local feature output vector set of the main branch network (for example, Figure 6 The first local feature output vector set H1 of the first main branch network and the first local feature output vector set H of the nth main branch network n ) corresponding to the sample person image; wherein the dimensions of the local feature output vectors of different main branch networks are kept consistent (for example, Figure 6 The first local feature output vector set H1 and the first local feature output vector set H n The feature vector dimension is kept at 1*N).

[0123] Step S330, calling each auxiliary branch network in the human body detection model to be trained to extract a local feature atlas matching the responsible preset human body part from the global feature atlas according to the human body part segmentation results of the character image training sample set, and performing local feature extraction on the local feature atlas to obtain a second local feature output vector set of each auxiliary branch network.

[0124] In this embodiment, the preset human body part composition involved in the human body part segmentation result is consistent with the preset human body part composition involved in the human body detection model to be trained, and each auxiliary branch network in the human body detection model to be trained is solely responsible for the local feature extraction operation of one preset human body part among the multiple preset human body parts.

[0125] When the skeleton network extracts the global feature map of any sample person image, it will synchronously input the global feature map of the sample person image to each auxiliary branch network. At this time, each auxiliary branch network will directly filter out the local feature map containing only the relevant features of the preset human body part from the received global feature map according to the actual distribution data of the preset human body part in the sample person image. At this time, the filtered local feature map is a component of the local feature map set corresponding to the auxiliary branch network. Then, the auxiliary branch network will directly extract local features from the filtered local feature map to directly obtain the second local feature output vector set of the auxiliary branch network (for example, Figure 6 The second local feature output vector set G1 of the first auxiliary branch network and the second local feature output vector set G of the nth auxiliary branch network n ) corresponding to the sample person image; wherein the dimensions of the local feature output vectors of different auxiliary branch networks are kept consistent (for example, Figure 6 The second local feature output vector set G1 and the second local feature output vector set G n Both maintain a 1*N feature vector dimension), and also maintain the same local feature output vector dimension as the corresponding main branch network (for example, Figure 1 All second local feature output vector sets and all first local feature output vector sets in the maintain a feature vector dimension of 1*N).

[0126] Step S340: For each main branch network, perform network optimization on the main branch network according to the second local feature output vector set and the first local feature output vector set corresponding to the main branch network.

[0127] In this embodiment, since the feature map input by any auxiliary branch network in the human body detection model to be trained when performing the local feature extraction operation only contains local features, the feature output vector extracted by any auxiliary branch network also only involves local features, the second local feature output vector set of the auxiliary branch network can be used as reference data for the network optimization of the corresponding main branch network, and the convolution layer weight parameters of the main branch network are dynamically adjusted by ensuring that the first local feature output vector set of the main branch network and the second local feature output vector set of the corresponding auxiliary branch network meet the pre-configured feature output vector difference constraint, thereby completing the network optimization operation of the main branch network. Among them, the feature output vector difference constraint can be described as "minimizing the feature output vector difference between the local feature output vectors of the kth main branch network and the kth auxiliary branch network in the human body detection model to be trained for the same sample person image", and the specific calculation formula of the feature output vector difference can be obtained by transforming the actual calculation formula of vector similarity such as cosine similarity, Jaccard similarity, Euclidean distance and Manhattan distance.

[0128] Specifically, the step of “for each main branch network, optimizing the main branch network according to the second local feature output vector set and the first local feature output vector set corresponding to the main branch network” may include:

[0129] For each sample person image in the person image training sample set, determine a second local feature output vector and a first local feature output vector that match the sample person image from the second local feature output vector set and the first local feature output vector set corresponding to the main branch network;

[0130] Calculating a feature output vector difference between a second local feature output vector and a first local feature output vector that match the sample person image;

[0131] With the purpose of minimizing the difference of the feature output vector, the convolutional layer weight parameters of the main branch network are optimized.

[0132] In one implementation of this embodiment, the specific calculation formula of the feature output vector difference is obtained by borrowing the actual calculation formula of cosine similarity, and the feature output vector difference of the kth main branch network and the kth auxiliary branch network in the human body detection model to be trained for the same sample person image is calculated using the following formula:

[0133]

[0134] Among them, C k It is used to represent the difference between the feature output vectors of the kth main branch network and the kth auxiliary branch network for the same sample person image, hk The first local feature output vector of the kth main branch network for the sample person image, g k The second local feature output vector of the kth auxiliary branch network for the sample person image. That is, it is used to represent the cosine similarity between the first local feature output vector of the kth main branch network and the second local feature output vector of the kth auxiliary branch network.

[0135] Step S360, detecting whether the model performance of the human body detection model to be trained reaches the expected performance state.

[0136] In this embodiment, after the optimization operation of each main branch network in the human body detection model to be trained is completed, it can be confirmed whether the model performance of the human body detection model to be trained at that time reaches the expected performance state by detecting whether all loss functions of the human body detection model to be trained at that time (including the loss functions of each main branch network, the loss functions of each auxiliary branch network, and the overall loss function of the human body detection model to be trained) have reached the minimum state. Then, when it is confirmed that the model performance of the human body detection model to be trained at that time has not reached the expected performance state, jump to the above-mentioned step S320 to continue execution until it is finally confirmed that the model performance of the human body detection model to be trained at that time reaches the expected performance state, thereby completing the training of the human body detection model to be trained. At this time, the computer device 10 will correspondingly execute step S370. Among them, the aforementioned loss function can be, but is not limited to, at least one of the conventional loss functions such as mean square error loss function, L2 loss function, L1 loss function, Smooth L1 loss function, huber loss function, relative entropy loss function, cross entropy loss function, softmax loss function, Focal loss function, etc.

[0137] Step S370, pruning all auxiliary branch networks in the trained human body detection model to be trained to obtain a human body part feature detection model.

[0138] In this embodiment, the present application trains the auxiliary branch network by obtaining local feature maps of different human body parts through human body segmentation. Even if there are incomplete human body images in the character image training sample set, it can be ensured that the training of the auxiliary branch network will not force the specific local areas of the incomplete human body images to be similar to the specific local areas of the complete human body images, thereby ensuring that the main branch network trained by borrowing the auxiliary branch network will not have the problem of misalignment of training features during the training process. Therefore, when the training of the human body detection model to be trained is completed, each main branch network in the human body detection model to be trained has the ability to extract local feature vectors that match the preset human body parts it is responsible for from the global feature map. At this time, all auxiliary branch networks in the human body detection model to be trained can be network pruned to obtain the final human body part feature detection model.

[0139] Therefore, the present application can ensure that the trained human body part feature detection model can realize the local feature extraction function of different human body parts for any character image by executing the above steps S310 to S370, and effectively solve the problem of training feature misalignment in the model training process.

[0140] In the embodiments provided in the present application, it should be understood that the disclosed devices and methods can also be implemented in other ways. The device embodiments described above are merely schematic, for example, the flowcharts and block diagrams in the accompanying drawings show the possible architecture, functions and operations of the devices, methods and computer program products according to the embodiments of the present application. In this regard, each box in the flowchart or block diagram can represent a module, a program segment or a part of a code, and the module, a program segment or a part of a code contains one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or the flowchart, and the combination of boxes in the block diagram and / or the flowchart can be implemented with a dedicated hardware-based system that performs a specified function or action, or can be implemented with a combination of dedicated hardware and computer instructions.

[0141] In addition, the functional modules in the various embodiments of the present application may be integrated together to form an independent part, or each module may exist separately, or two or more modules may be integrated to form an independent part.

[0142] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a readable storage medium. Based on this understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a readable storage medium, including several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in each embodiment of the present application. The aforementioned readable storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0143] The above are only various implementations of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A target person tracking method, characterized in that: The method comprises: Acquire a real-time image of the person to be detected and a reference image of the target person; Calling a pre-trained human body part feature detection model to extract a first local feature vector of at least one preset human body part actually existing in the real-time human image, and a second local feature vector of at least one preset human body part actually existing in the reference human image; Detecting whether the real-time character image and the reference character image have at least one target human body part in common based on the extracted first local feature vector and the second local feature vector; If it is detected that the real-time person image and the reference person image do not have at least one target human body part in common, it is determined that the person to be detected is not the target person, otherwise it is detected whether the feature similarity between the first target feature vector and the second target feature vector exceeds a preset similarity threshold, wherein the first target feature vector is obtained by merging the first local feature vectors of all target human body parts, and the second target feature vector is obtained by merging the second local feature vectors of all target human body parts; When it is detected that the corresponding feature similarity exceeds a preset similarity threshold, it is determined that the person to be detected is the target person, otherwise it is determined that the person to be detected is not the target person.

2. The method according to claim 1, characterized in that The human body part feature detection model includes a backbone network and a plurality of main branch networks, wherein each main branch network is independently responsible for a local feature extraction operation of a preset human body part, and the step of calling the pre-trained human body part feature detection model to extract a first local feature vector of at least one preset human body part actually existing in the real-time human image, and a second local feature vector of at least one preset human body part actually existing in the reference human image, comprises: For each person image in the real-time person image and the reference person image, calling the backbone network to extract image features of the person image to obtain a global feature map of the person image; The global feature map of the person image is input to all main branch networks at the same time, and each main branch network is called to extract the actual local feature vector matching the responsible preset human body part from the input global feature map; For each actual local feature vector of a preset human body part corresponding to the person image, detecting whether the actual local feature vector of the preset human body part is valid; When it is detected that the actual local feature vector of the preset human body part is valid, the actual local feature vector of the preset human body part is used as a target local feature vector actually existing in the character image, wherein the target local feature vector is a first local feature vector actually existing in the real-time character image, or the target local feature vector is a second local feature vector actually existing in the reference character image.

3. The method according to claim 2, characterized in that The step of detecting whether the actual local feature vector of each preset human body part corresponding to the character image is valid comprises: Calculating the vector modulus of the actual local feature vector of the preset human body part; Detecting whether the calculated vector modulus exceeds a preset modulus threshold value that matches the preset human body part, wherein different preset human body parts correspond to different preset modulus threshold values; If it is detected that the vector modulus exceeds a preset modulus threshold matching the preset human body part, the actual local feature vector of the preset human body part is determined to be valid; otherwise, the actual local feature vector of the preset human body part is determined to be invalid.

4. The method according to claim 1, characterized in that: The step of detecting whether the real-time character image and the reference character image have at least one target human body part in common based on the extracted first local feature vector and second local feature vector comprises: Performing intersection processing on the preset human body parts corresponding to all the first local feature vectors and the preset human body parts corresponding to all the second local feature vectors to obtain a corresponding human body part result set; When the human body part result set is an empty set, determining that the real-time human image and the reference human image do not have at least one target human body part in common; When the human body part result set is a non-empty set, it is determined that the real-time human image and the reference human image have at least one target human body part in common, and all preset human body parts in the human body part result set are respectively regarded as a target human body part.

5. The method according to any one of claims 1 to 4, characterized in that: The method further comprises: Acquire a human image training sample set for a human body detection model to be trained, and human body part segmentation results of the human image training sample set that match multiple preset human body parts; Calling the backbone network in the human body detection model to be trained to extract image features from the human image training sample set to obtain a corresponding global feature atlas; Calling each main branch network in the human body detection model to be trained to extract local features from the global feature atlas, respectively, to obtain a first local feature output vector set of each main branch network; Calling each auxiliary branch network in the human body detection model to be trained to extract a local feature atlas matching the responsible preset human body part from the global feature atlas according to the human body part segmentation result of the human image training sample set, and performing local feature extraction on the local feature atlas to obtain a second local feature output vector set of each auxiliary branch network, wherein each auxiliary branch network in the human body detection model to be trained is solely responsible for the local feature extraction operation of one preset human body part among the multiple preset human body parts, and each auxiliary branch network corresponds to a main branch network and shares convolution layer weight parameters with the main branch network; For each main branch network, the main branch network is optimized according to the second local feature output vector set and the first local feature output vector set corresponding to the main branch network, and then the step of calling the backbone network to extract image features from the person image training sample set is continued until the model performance of the to-be-trained human body detection model reaches the expected performance state; All auxiliary branch networks in the trained human body detection model to be trained are pruned to obtain the human body part feature detection model.

6. The method according to claim 5, characterized in that The step of obtaining the human body part segmentation results matching a plurality of preset human body parts of the human image training sample set comprises: For each sample person image in the person image training sample set, calling a human body part segmentation model to perform part distribution recognition in the sample person image according to the multiple preset human body parts, and obtaining actual distribution data of each of the multiple preset human body parts in the sample person image; The actual distribution data corresponding to the plurality of preset human body parts of all sample human images in the human image training sample set are sorted to obtain the human body part segmentation result of the human image training sample set.

7. The method according to claim 5, characterized in that For each main branch network, the step of performing network optimization on the main branch network according to the second local feature output vector set and the first local feature output vector set corresponding to the main branch network includes: For each sample person image in the person image training sample set, determine a second local feature output vector and a first local feature output vector that match the sample person image from the second local feature output vector set and the first local feature output vector set corresponding to the main branch network; Calculating a feature output vector difference between a second local feature output vector and a first local feature output vector that match the sample person image; With the purpose of minimizing the difference of the feature output vector, the convolutional layer weight parameters of the main branch network are optimized.

8. The method according to claim 7, characterized in that The difference of the feature output vectors of the kth main branch network and the kth auxiliary branch network in the human detection model to be trained for the same sample person image is calculated using the following formula: Among them, C k It is used to represent the difference between the feature output vectors of the kth main branch network and the kth auxiliary branch network for the same sample person image, h k The first local feature output vector of the kth main branch network for the sample person image, g k The second local feature output vector of the kth auxiliary branch network for the sample person image.

9. A computer device, characterized in that: The invention comprises a processor and a memory, wherein the memory stores a computer program executable by the processor, and the processor can execute the computer program to implement the target person tracking method described in any one of claims 1 to 8.

10. A readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a computer device, the target person tracking method described in any one of claims 1 to 8 is implemented.

Images