Multi-target tracking method, device and storage medium based on appearance feature quality screening

Abstract

The present invention discloses a multi-target tracking method, device and storage medium based on appearance feature quality screening, which extracts appearance features of detection targets by constructing a lightweight appearance feature network; performs quality assessment on the appearance features to generate quality scores of corresponding detection targets; divides the detection targets into different quality levels according to the quality scores, and first performs inter-frame association operations on a first target image set and target position information of corresponding frames according to the order of quality, so as to obtain tracking trajectories of the detection targets of the first target image set; then performs inter-frame association operations on a second target image set and target position information of corresponding frames, so as to obtain tracking trajectories of the detection targets of the second target image set; merges the tracking trajectories of the first target image set and the tracking trajectories of the second target image set, so as to obtain the final tracking trajectories of the detection targets; thereby maintaining high tracking accuracy while also improving tracking efficiency.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to a multi-target tracking method based on appearance feature quality screening, a device applying the method, and a computer-readable storage medium. Background Art

[0002] Multi-Object Tracking (MOT) is an important research direction in the field of computer vision, which aims to accurately track and identify multiple targets in a scene at the same time. With the continuous development of computer vision and deep learning technology, the application prospects of multi-object tracking algorithms in various fields are becoming more and more broad. In the field of autonomous driving, the MOT algorithm can accurately track road users such as pedestrians and vehicles, providing strong support for safe driving and traffic management. In the field of video surveillance, the MOT algorithm can continuously track and analyze the trajectories of multiple targets in the monitoring scene, providing an important basis for security prevention and incident investigation. In addition, the MOT algorithm can also be applied to multiple fields such as human-computer interaction, behavior analysis, sports video analysis, etc., with broad application prospects and important research value.

[0003] The multi-target tracking algorithm mainly combines target detection, feature extraction and data association technology to continuously track and reconstruct the trajectory of multiple targets in the video sequence. It first uses the target detection algorithm to identify the target in each frame, then extracts the target's feature information, and matches the targets in different frames through the data association algorithm to construct the complete motion trajectory of each target.

[0004] Compared with traditional single target tracking, multi-target tracking needs to process more targets, and the accuracy of the algorithm is affected by many factors such as occluded objects, scale changes, appearance changes, motion blur, etc. In addition, the high computational complexity also restricts the application of multi-target tracking algorithms on resource-constrained devices. Summary of the invention

[0005] The main purpose of the present invention is to provide a multi-target tracking method, device and storage medium based on appearance feature quality screening, aiming to solve the technical problem of how to balance the accuracy and computational efficiency of the multi-target tracking algorithm.

[0006] To achieve the above object, the present invention provides a multi-target tracking method based on appearance feature quality screening, which comprises the following steps:

[0007] A lightweight appearance feature network is constructed by using a network reparameterization technique, and the appearance features of the detection target are extracted through the appearance feature network;

[0008] Performing quality assessment on the appearance features to generate a quality score of the corresponding detection target;

[0009] According to the quality score, the detection targets are divided into different quality levels, target images with quality scores greater than or equal to a preset threshold are screened to obtain a first target image set, and target images with quality scores less than the preset threshold are screened to obtain a second target image set;

[0010] Use the trained target detector to detect the input video or image set frame by frame to obtain the target position information in each frame;

[0011] According to the order of quality, firstly, the first target image set and the target position information of each corresponding frame are subjected to an inter-frame association operation to obtain the tracking trajectory of the detection target of the first target image set; then, based on the matching result of the first target image set, the second target image set and the target position information of each corresponding frame are subjected to an inter-frame association operation to obtain the tracking trajectory of the detection target of the second target image set;

[0012] The tracking trajectory of the first target image set is combined with the tracking trajectory of the second target image set to obtain a final tracking trajectory of the detected target.

[0013] Preferably, the appearance feature network includes 5 stages, the first stage includes 1 convolutional network module, and the second to fifth stages respectively include 2 repeatedly stacked convolutional network modules to form a 9-layer convolutional network structure; a quality assessment module for performing quality assessment on the appearance features is also provided behind the output layer of the fifth stage, and the quality assessment module sequentially includes a first fully connected layer with an output length of 128 dimensions, a second fully connected layer with an output length of 1, and a sigmoid activation function.

[0014] Preferably, the calculation formula of the mass fraction is:

[0015]

[0016] in, is the weight matrix of the first fully connected layer, is the bias term of the first fully connected layer; is the weight matrix of the second fully connected layer, is the bias term of the second fully connected layer; is the sigmoid activation function; is the weight matrix of the output layer of the 5th stage of the appearance feature network.

[0017] Preferably, the loss function of the quality assessment module is:

[0018]

[0019] in, Represents the total number of samples that meet the conditions in the training data set, and They represent the appearance feature vector of the i-th target image and the appearance feature vector of the j-th target image, respectively. Represents the dimension index of the appearance feature vector of the detected target, and Represent the appearance feature vectors and No. Dimensions, and They represent the quality score of the i-th target image and the quality score of the j-th target image respectively.

[0020] Preferably, the target detector is a YOLOX detector for the input video sequence. Detect the target frame by frame, obtain the position information of the detected target in each frame, and obtain the target position set; based on the position information, intercept the corresponding target image in the frame image; for any video, Frame image, get the target image set ,in Indicates the video Frame image, Indicates the first target images; extracting features from the target image set using an appearance feature network to obtain a set of appearance feature vectors of the target images ; Perform quality assessment on appearance features and obtain the quality score set of the target image .

[0021] Preferably, the quality score set of the target image obtained according to the t-th frame image is , set the threshold to associate the target image set The target images with quality scores greater than or equal to the preset threshold are divided into two groups and are defined as , represents the first target image set; the remaining target images with quality scores less than the preset threshold are defined as , represents the second target image set.

[0022] Furthermore, define The trajectory set at the frame time is , Indicates the first target tracks; for each target track , and its corresponding appearance feature vector is expressed as ; Based on the first similarity calculation formula, the first target image set The current trajectory set is Performing target association operation, the first similarity calculation formula is:

[0023]

[0024] in, represents the first target image trajectory matching similarity, represents the cosine similarity, Represents the target image Bounding box and target track to be matched The intersection of bounding boxes; according to the first target image set The set of trajectories to be matched The similarity between them is measured by using the Hungarian algorithm to match and update the existing trajectories. ; After association, The remaining target image in is defined as ,Will The unmatched trajectory in is defined as .

[0025] Further, based on the second similarity calculation formula, the unmatched trajectory set With the second target image set Target association The second similarity calculation formula is:

[0026]

[0027] in represents the second target image trajectory matching similarity, for The quality score, represents the cosine similarity, Represents the target image Bounding box and target track to be matched The intersection of bounding boxes; according to the second target image set The set of trajectories to be matched is The similarity between them is measured by using the Hungarian algorithm to match and update the existing trajectories and merge the results into After association, The unmatched trajectory is defined as ; Delete unmatched ,against Initialize the new trajectory from the target image in and update it to middle.

[0028] In addition, to achieve the above-mentioned purpose, the present invention also provides a device, which includes a memory, a processor, and a multi-target tracking program stored in the memory and executable on the processor, and when the multi-target tracking program is executed by the processor, the steps of the multi-target tracking method based on appearance feature quality screening as described above are implemented.

[0029] In addition, to achieve the above-mentioned purpose, the present invention also provides a computer-readable storage medium, on which a multi-target tracking program is stored. When the multi-target tracking program is executed by a processor, the steps of the multi-target tracking method based on appearance feature quality screening as described above are implemented.

[0030] The beneficial effects of the present invention are:

[0031] (1) By using network reparameterization technology to construct a lightweight appearance feature network, the present invention can effectively reduce the computational complexity of the model without sacrificing too much accuracy, making the extraction of appearance features more efficient. At the same time, the lightweight network design facilitates deployment and operation on resource-constrained devices.

[0032] (2) The present invention improves the correlation performance of detection frames between adjacent frames by designing a lightweight appearance feature network, thereby alleviating the impact of factors such as object occlusion, scale change, appearance change, motion blur, etc. on the multi-target tracking task.

[0033] (3) The present invention introduces a feature-based quality screening module, which divides detection targets of different qualities and performs inter-frame correlation operations to improve the robustness and accuracy of the pedestrian multi-target tracking algorithm.

[0034] (4) The present invention evaluates the quality of appearance features and classifies the quality levels accordingly, so that the algorithm can give priority to processing target images with higher quality, significantly improving the accuracy and robustness of tracking, because high-quality target images usually contain clearer feature information and are easier to accurately associate between frames.

[0035] (5) The detected targets are divided into different sets according to their quality scores, and high-quality target images are processed first. This strategy effectively optimizes the allocation of computing resources. For target images with lower quality, although the generation of their tracking trajectories is delayed, computing resources are saved overall, allowing the algorithm to respond and process more quickly overall. By processing high-quality target images first and then low-quality target images, the present invention improves tracking efficiency while maintaining high tracking accuracy.

[0036] (6) The present invention divides the appearance feature network into five stages, each of which gradually extracts deeper features. The first stage uses a single convolutional network module as a starting point, while the second to fifth stages deepen the network by repeatedly stacking two convolutional network modules. This design helps to gradually capture feature representations from low-level to high-level; the entire network consists of a 9-layer convolutional network. This moderate depth design provides sufficient capacity to learn complex appearance features while maintaining computational efficiency.

[0037] (7) The present invention also introduces a quality score when calculating the similarity of low-quality target images; by multiplying the quality score with the cosine similarity, a new similarity measurement value is obtained; this value not only takes into account the similarity of the image content (measured by the cosine similarity) but also takes into account the quality of the image (measured by the quality score); therefore, even if two images are very similar in content, if the quality of one of the images is very low, then their comprehensive similarity will be reduced accordingly, so that the similarity in the low-quality target image set can be more comprehensively evaluated, providing a more accurate and reliable basis for related applications.

[0038] (8) The multi-target tracking algorithm based on appearance feature quality screening proposed in the present invention can be efficiently deployed in real time on various computing-constrained terminal devices. In practical application scenarios such as autonomous driving, video surveillance, and human-computer interaction, the method of the present invention can significantly improve the performance of the tracking system and provide more accurate and reliable data support for subsequent decision analysis and event processing. In particular, under complex and changing environmental conditions, such as occlusion, scale changes, and appearance changes, the method of the present invention exhibits stronger adaptability and stability. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the following will be clearly and completely described in combination with the specific technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0040] A multi-target tracking method based on appearance feature quality screening of the present invention comprises the following steps:

[0041] A lightweight appearance feature network is constructed by using a network reparameterization technique, and the appearance features of the detection target are extracted through the appearance feature network;

[0042] Performing quality assessment on the appearance features to generate a quality score of the corresponding detection target;

[0043] According to the quality score, the detection targets are divided into different quality levels, target images with quality scores greater than or equal to a preset threshold are screened to obtain a first target image set, and target images with quality scores less than the preset threshold are screened to obtain a second target image set;

[0044] Use the trained target detector to detect the input video or image set frame by frame to obtain the target position information in each frame;

[0045] According to the order of quality, firstly, the first target image set and the target position information of each corresponding frame are subjected to an inter-frame association operation to obtain the tracking trajectory of the detection target of the first target image set; then, based on the matching result of the first target image set, the second target image set and the target position information of each corresponding frame are subjected to an inter-frame association operation to obtain the tracking trajectory of the detection target of the second target image set;

[0046] The tracking trajectory of the first target image set is combined with the tracking trajectory of the second target image set to obtain a final tracking trajectory of the detected target.

[0047] The existing technology mainly uses DeepSort (Deep Simple Online and Realtime Tracking, a multi-target tracking algorithm based on deep learning). DeepSort introduces a target re-identification network based on deep features to associate targets between frames. However, due to the large amount of computation required by the target re-identification network, DeepSort is difficult to deploy in real time on low-computing power chips such as mobile terminals.

[0048] In this embodiment, a lightweight appearance feature network is adopted, as shown in Table 1, the appearance feature network includes 5 stages, the first stage includes 1 convolutional network module, the second stage to the fifth stage respectively include 2 repeatedly stacked convolutional network modules, forming a 9-layer convolutional network structure; the output layer of the fifth stage is also provided with a quality assessment module for performing quality assessment on the appearance features, and the quality assessment module sequentially includes a first fully connected layer with an output length of 128 dimensions, a second fully connected layer with an output length of 1, and a sigmoid activation function.

[0049]

[0050] Table 1 - Schematic diagram of the structure of the lightweight appearance feature network of the present invention

[0051] In Table 1, the feature output size is expressed as width high Channel, module composition Represents repeated stacking of n convolutional network modules A.

[0052] In this embodiment, the calculation formula of the mass fraction is:

[0053]

[0054] in, is the weight matrix of the first fully connected layer, is the bias term of the first fully connected layer; is the weight matrix of the second fully connected layer, is the bias term of the second fully connected layer; is the sigmoid activation function; is the weight matrix of the output layer of the 5th stage of the appearance feature network.

[0055] The expression of the sigmoid activation function is:

[0056]

[0057] Given an input image and , and their sample category labels are and ,Using the above lightweight appearance feature network, the appearance feature vectors of the detected target can be defined as and , the quality scores can be obtained by using the quality assessment module, which are defined as and .

[0058] Therefore, the loss function of the quality assessment module is:

[0059]

[0060] in, Represents the total number of samples that meet the conditions in the training data set, and They represent the appearance feature vector of the i-th target image and the appearance feature vector of the j-th target image, respectively. Represents the dimension index of the appearance feature vector of the detected target, and Represent the appearance feature vectors and No. Dimensions, and They represent the quality score of the i-th target image and the quality score of the j-th target image respectively.

[0061] Based on the loss function, all parameters of the lightweight appearance feature network are fixed, and the parameters in the quality assessment module are trained using stochastic gradient descent. In this embodiment, a private data set is used as the training data for the appearance feature network, and the data set contains 28,358 pedestrian targets and 242,584 target images. The trained quality screening module based on appearance features is applied to the multi-target tracking method proposed in the present invention, which can effectively divide the detection targets of different qualities, and perform inter-frame target association operations in batches according to the quality.

[0062] Traditional multi-target tracking algorithms will simply discard low-quality detection targets caused by factors such as occluded objects, motion blur, and illumination changes, resulting in a large number of missed detections and fragmented trajectories. The quality screening module based on appearance features in this embodiment can divide the targets to be associated into a set of high-quality target images and a set of low-quality target images according to the quality scores, and apply different inter-frame association operations to the two types of quality target image data, effectively processing low-quality target images (such as pedestrian images) to ensure the integrity of the target tracking trajectory.

[0063] In this embodiment, given a video sequence to be processed , define The trajectory set at the frame time is , Indicates the first For each target trajectory , the appearance feature of the detected target of its trajectory is expressed as .

[0064] The target detector uses the YOLOX detector to detect the input video sequence. Detect the target frame by frame, obtain the position information of the detected target in each frame, and obtain the target position set; based on the position information, intercept the corresponding target image in the frame image; for any video, Frame image, get the target image set ,in Indicates the video Frame image, Indicates the first target images; extracting features from the target image set using an appearance feature network to obtain a set of appearance feature vectors of the target images ; Perform quality assessment on appearance features and obtain the quality score set of the target image .

[0065] The YOLOX detector is a high-performance anchor-free target detector. Its structure mainly consists of four parts: input end, backbone network (Backbone), neck network (Neck) and prediction head (Head).

[0066] In this embodiment, the input end of the YOLOX detector adopts the Mosaic data enhancement method to generate new training samples by cropping and splicing four pictures, thereby increasing the diversity of samples; the backbone network adopts CSPDarknet, and uses a convolutional neural network for feature extraction, and gradually extracts deep features in the image by stacking multiple convolutional layers and residual blocks; the neck network adopts structures such as feature pyramid network (FPN) or path aggregation network (PAFPN), and fuses features of different scales extracted by the backbone network through upsampling and downsampling operations; the prediction head adopts an anchor-free frame detection mechanism to directly predict the coordinates and width and height of the center point of the object, avoiding the limitations and complexity brought by the anchor frame design, and can more flexibly handle targets of different shapes and sizes.

[0067] In traditional multi-target tracking methods, the intersection over union (IoU) of target detection boxes in different frames is usually used for target association. When multiple target detection boxes overlap, misassociation is prone to occur.

[0068] In this embodiment, the quality score set of the target image obtained according to the t-th frame image , set the threshold to associate the target image set The target images with quality scores greater than or equal to the preset threshold are divided into two groups and are defined as , represents the first target image set; the remaining target images with quality scores less than the preset threshold are defined as , represents the second target image set. And, based on the similarity calculation, the target image set and the target position set are associated with each other to obtain the tracking trajectory of the detected target; wherein, the first target image set is associated with each other based on the first similarity calculation formula, and the second target image set is associated with each other based on the second similarity calculation formula.

[0069] In this embodiment, for the first The quality score set of the target image obtained by the frame image , set the threshold to associate the current frame target image set The detection results with scores higher than the threshold are defined as , represents the first target image set (high-quality target image set), and the remaining detection results are defined as , represents the second target image set (low-quality target image set). For example, the threshold is defined as 0.3, that is, target images with a quality score higher than 0.3 enter the first target image set, and target images with a quality score lower than 0.3 enter the second target image set. In addition, a second threshold may be set for the first target image set as needed to further divide it into a third target image set and a fourth target image set; and / or a third threshold may be set for the second target image set to further divide it into a fifth target image set and a sixth target image set; and so on.

[0070] In this embodiment, define The trajectory set at the frame time is , Indicates the first target tracks; for each target track , and its corresponding appearance feature vector is expressed as ; Based on the first similarity calculation formula, the first target image set The current trajectory set is Performing target association operation, the first similarity calculation formula is:

[0071]

[0072] in, represents the first target image trajectory matching similarity, represents the cosine similarity, Represents the target image Bounding box and target track to be matched The intersection of bounding boxes; according to the first target image set The set of trajectories to be matched The similarity between them is measured by using the Hungarian algorithm to match and update the existing trajectories. .

[0073] After association, The remaining targets in are defined as ,Will The unmatched tracking trajectory is defined as .

[0074] In this embodiment, the unmatched trajectory set is further calculated based on the second similarity calculation formula. With the second target image set Target association The second similarity calculation formula is:

[0075]

[0076] in represents the second target image trajectory matching similarity, for The quality score, represents the cosine similarity, Represents the target image Bounding box and target track to be matched The intersection of bounding boxes; according to the second target image set The set of trajectories to be matched is The similarity between them is measured by using the Hungarian algorithm to match and update the existing trajectories and merge the results into middle.

[0077] After association, The unmatched tracking trajectory in is defined as .

[0078] Finally, delete the unmatched ,against Initialize the new trajectory with the target image in and update it into middle.

[0079] After testing, the multi-target tracking algorithm of the present invention can achieve 72.3% multiple object tracking accuracy (Multiple Object Tracking Accuracy, MOTA) even with a lightweight network.

[0080] In addition, the multi-target tracking method based on appearance feature quality screening proposed in the present invention can be efficiently and real-time deployed on computing-constrained embedded terminal devices. For example, the algorithm can achieve a computing performance of 35 frames per second (35 fps) on the HiSilicon chip Hi3516CV500.

[0081] In addition, the present invention also provides a multi-target tracking device, which can execute the technical solution of the method embodiment. Its implementation principle and technical effect are similar. For details, please refer to the relevant records in the above method embodiment, which will not be repeated here.

[0082] The device includes: a device with a camera function such as a mobile phone, a digital camera or a tablet computer, or a device with an image processing function, or a device with an image display function. The device may include components such as a memory, a processor, an input unit, a display unit, and a power supply.

[0083] Among them, the memory can be used to store software programs and modules, and the processor executes various functional applications and data processing by running the software programs and modules stored in the memory. The memory may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function (such as an image playback function, etc.), etc.; the data storage area may store data created according to the use of the device, etc. In addition, the memory may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other volatile solid-state storage devices. Accordingly, the memory may also include a memory controller to provide the processor and the input unit with access to the memory.

[0084] The input unit can be used to receive input digital or character or image information, and generate keyboard, mouse, joystick, optical or trackball signal input related to user settings and function control. Specifically, the input unit of this embodiment includes not only a camera, but also a touch-sensitive surface (such as a touch display) and other input devices.

[0085] The display unit can be used to display information input by the user or information provided to the user and various graphical user interfaces of the device, which can be composed of graphics, text, icons, videos and any combination thereof. The display unit may include a display panel, and optionally, the display panel may be configured in the form of LCD (Liquid Crystal Display), OLED (Organic Light-Emitting Diode), etc. Further, the touch-sensitive surface may cover the display panel, and when the touch-sensitive surface detects a touch operation on or near it, it is transmitted to the processor to determine the type of touch event, and then the processor provides a corresponding visual output on the display panel according to the type of touch event.

[0086] The embodiment of the present invention further provides a computer-readable storage medium, which may be a computer-readable storage medium contained in the memory in the above embodiment; or a computer-readable storage medium that exists independently and is not assembled into a device. The computer-readable storage medium stores at least one instruction, which is loaded and executed by a processor to implement a multi-target tracking method based on appearance feature quality screening. The computer-readable storage medium may be a read-only memory, a disk or an optical disk, etc.

[0087] It should be noted that each embodiment in this specification is described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same or similar parts between the embodiments can be referred to each other. For the device embodiment and the storage medium embodiment, since they are basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0088] Furthermore, in this document, the terms "comprises," "comprising," or any other variation thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or device that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article, or device that includes the element.

[0089] The above description shows and describes the preferred embodiments of the present invention. It should be understood that the present invention is not limited to the form disclosed herein, and should not be regarded as excluding other embodiments, but can be used in various other combinations, modifications and environments, and can be modified within the scope of the invention, through the above teachings or the technology or knowledge of the relevant field. The changes and modifications made by those skilled in the art shall not depart from the spirit and scope of the present invention, and shall be within the scope of protection of the claims attached to the present invention.

Claims

1. A multi-target tracking method based on appearance feature quality screening, characterized in that: The following steps are involved: A lightweight appearance feature network is constructed by using a network reparameterization technique, and the appearance features of the detection target are extracted through the appearance feature network; Performing quality assessment on the appearance features to generate a quality score of the corresponding detection target; According to the quality score, the detection targets are divided into different quality levels, target images with quality scores greater than or equal to a preset threshold are screened to obtain a first target image set, and target images with quality scores less than the preset threshold are screened to obtain a second target image set; Use the trained target detector to detect the input video or image set frame by frame to obtain the target position information in each frame; According to the order of quality, firstly, the first target image set and the target position information of each corresponding frame are subjected to an inter-frame association operation to obtain the tracking trajectory of the detection target of the first target image set; then, based on the matching result of the first target image set, the second target image set and the target position information of each corresponding frame are subjected to an inter-frame association operation to obtain the tracking trajectory of the detection target of the second target image set; Merging the tracking trajectory of the first target image set with the tracking trajectory of the second target image set to obtain a final tracking trajectory of the detected target; The appearance feature network includes 5 stages, the first stage includes 1 convolutional network module, and the second to fifth stages respectively include 2 repeatedly stacked convolutional network modules to form a 9-layer convolutional network structure; a quality assessment module for performing quality assessment on the appearance features is also provided behind the output layer of the fifth stage, and the quality assessment module sequentially includes a first fully connected layer with an output length of 128 dimensions, a second fully connected layer with an output length of 1, and a sigmoid activation function; The target detector uses the YOLOX detector to detect the input video sequence. Detect the target frame by frame, obtain the position information of the detected target in each frame, and obtain the target position set; based on the position information, intercept the corresponding target image in the frame image; for any video, Frame image, get the target image set ,i=0,1,…n; where Indicates the video Frame image, Indicates the first target images; extracting features from the target image set using an appearance feature network to obtain a set of appearance feature vectors of the target images ,i=0,1,…n; perform quality assessment on appearance features and obtain the quality score set of the target image , i=0,1,…n.

2. The multi-target tracking method based on appearance feature quality screening according to claim 1, characterized in that: The calculation formula of the mass fraction is: in, is the weight matrix of the first fully connected layer, is the bias term of the first fully connected layer; is the weight matrix of the second fully connected layer, is the bias term of the second fully connected layer; is the sigmoid activation function; is the weight matrix of the output layer of the 5th stage of the appearance feature network.

3. The multi-target tracking method based on appearance feature quality screening according to claim 1, characterized in that: The loss function of the quality assessment module is: in, Represents the total number of samples that meet the conditions in the training data set, and They represent the appearance feature vector of the i-th target image and the appearance feature vector of the j-th target image, respectively. Represents the dimension index of the appearance feature vector of the detected target, and Represent the appearance feature vectors and No. Dimensions, and They represent the quality score of the i-th target image and the quality score of the j-th target image respectively.

4. The multi-target tracking method based on appearance feature quality screening according to claim 1, characterized in that: The quality score set of the target image obtained according to the t-th frame image , i=0,1,…n; Set the threshold to associate the target image set The target images with quality scores greater than or equal to the preset threshold are divided into two groups and are defined as , represents the first target image set; the remaining target images with quality scores less than the preset threshold are defined as , represents the second target image set.

5. The multi-target tracking method based on appearance feature quality screening according to claim 4, characterized in that: Definition The trajectory set at the frame time is , =0,1,…n; Indicates the first target tracks; for each target track , and its corresponding appearance feature vector is expressed as ; Based on the first similarity calculation formula, the first target image set The current trajectory set is Performing target association operation, the first similarity calculation formula is: in, represents the matching similarity between the first target image set and the current trajectory set, represents the cosine similarity, Represents the target image Bounding box and target track to be matched The intersection of bounding boxes; according to the first target image set With the current track collection The similarity between them is measured by using the Hungarian algorithm to match and update the existing trajectories. ; After association, The remaining target image in is defined as ,Will The unmatched trajectory in is defined as .

6. The multi-target tracking method based on appearance feature quality screening according to claim 5, characterized in that: Based on the second similarity calculation formula, the unmatched trajectory set With the second target image set Perform target association; the second similarity calculation formula is: in represents the matching similarity between the second target image set and the unmatched trajectory set, for The quality score, represents the cosine similarity, Represents the target image Bounding box and target track to be matched The intersection of bounding boxes; according to the second target image set The set of unmatched trajectories The similarity between them is measured by using the Hungarian algorithm to match and update the existing trajectories and merge the results into After association, The unmatched trajectory is defined as ; Delete unmatched ,against Initialize the new trajectory from the target image in and update it to middle.

7. A multi-target tracking device based on appearance feature quality screening, characterized in that: The device includes a memory, a processor, and a multi-target tracking program stored in the memory and executable on the processor. When the multi-target tracking program is executed by the processor, the steps of the multi-target tracking method based on appearance feature quality screening as described in any one of claims 1 to 6 are implemented.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a multi-target tracking program, which, when executed by a processor, implements the steps of the multi-target tracking method based on appearance feature quality screening as described in any one of claims 1 to 6.

Images