Surrounding target re-identification method, system and monitoring equipment
By using preset camera array and target recognition model in the circumferential target recognition method, combined with convolutional neural network and twin network, the problem of target recognition across lenses and cross-scene is solved, and fast and accurate target tracking and recognition is achieved. Improved the identification capability of the monitoring system.
Patent Information
- Application Number
- CN202410520765.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-04-28
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2044-04-28
AI Technical Summary
In the prior art, how to quickly correlate and accurately monitor the movement trajectory of targets in different cameras and different video clips is a technical problem that needs to be solved urgently, especially in target recognition and retrieval across lenses and scenes.
Using the circumferential target recognition method, the real-time image is obtained through the preset camera array and the pre-trained target recognition model is input, the target is identified and its associated image sequence is determined, and the target is tracked by similarity calculation and exposure order arrangement is used to track the target's movement trajectory. . The method includes a regular polygonal arrangement of the camera array and a clockwise and counterclockwise exposure sequence, combining the cascade of the convolutional neural network and the twin network to achieve target identification and tracking.
The image sequence analysis with correlation between time series and spatial sequence is realized, which can quickly and accurately identify and track targets, improve the target recognition and retrieval capabilities across cameras and scenes, and has 360-degree field of view and high reliability.
Smart Images

Figure CN118379662B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of target recognition, and in particular to a surround view target re-recognition method, system and monitoring equipment. Background Art
[0002] Object re-identification (Re-ID) is one of the research hotspots in the field of computer vision. Its research goal is to determine whether the target images captured by different cameras or in different video clips of the same camera are the same target. Object re-identification is also called cross-camera tracking, which can be regarded as an image retrieval task. It uses computer vision technology to retrieve whether there are targets of the same identity in images or video sequences. It mainly aims to solve the recognition and retrieval of targets across cameras and scenes.
[0003] In the existing technology, how to quickly correlate and analyze the recognition results of the same target in different cameras and different video clips and accurately monitor the target's movement trajectory is a technical problem that needs to be solved urgently. Therefore, it is necessary to study a target re-identification system to achieve cross-lens rapid tracking of the target object in the target scene. Summary of the invention
[0004] The present specification provides a surround target re-identification method, system and monitoring device to overcome at least one technical problem existing in the related art.
[0005] According to a first aspect of an embodiment of the present specification, a surround target re-identification method is provided, comprising: inputting a real-time image obtained from a preset camera array into a pre-trained target re-identification model, forming a target database with targets identified by the target re-identification model, the target re-identification model detecting the input image to identify the target in the image, and then determining an image sequence associated with the target according to the identified target; obtaining an initial image containing a target to be retrieved, inputting the initial image containing the target to be retrieved into the target re-identification model to obtain the target to be retrieved, using the target re-identification model to perform similarity calculation on the target to be retrieved in the initial image in a target database, determining associated targets of the target to be retrieved in the target database, and arranging the associated targets in descending order of similarity; based on the associated targets obtained, calling out an image sequence containing the associated targets captured by the camera array according to a preset exposure sequence; obtaining a movement trajectory of the target to be retrieved according to the timestamps of the images in the image sequence and the shooting positions of each camera in the corresponding exposure sequence, so as to realize the recognition and tracking of the target to be retrieved.
[0006] Optionally, the camera array includes a preset number of cameras arranged in a preset shape, wherein the preset number of cameras are evenly arranged along the circumference of a regular polygon to form a camera array, and the camera array exposes the cameras in sequence according to a set exposure order, and the exposure order includes a clockwise sequence and a counterclockwise sequence, and one of any two adjacent cameras is designated as a clockwise starting camera, and the other camera is designated as a counterclockwise starting camera. Starting from the clockwise starting camera, exposure is sequentially performed in a clockwise direction with one camera interval at a time to form a clockwise sequence, and starting from the counterclockwise starting camera, exposure is sequentially performed in a counterclockwise direction with one camera interval at a time to form a counterclockwise sequence. The clockwise sequence and the counterclockwise sequence start exposure at the same time or start exposure with a preset time difference.
[0007] Optionally, before the step of inputting the real-time images obtained from the preset camera array into the pre-trained target re-identification model, it also includes: labeling the images obtained from the camera array within a preset time length; using the images obtained by the camera array and the corresponding labeled images to form training data according to preset classifications, and training and obtaining the target re-identification model through the training data, wherein the preset classifications are related to the layout shape of the camera array and the set exposure sequence.
[0008] Optionally, the step of annotating the images acquired from the camera array within a preset time length includes three methods, wherein the first method is: adding an annotation box to the image through manual cropping; the second method is: processing the image through an algorithm model with a target detection function, and adding an annotation box to the image; the third method is: processing the image through an algorithm model with a target tracking function, and adding an annotation box to the image.
[0009] Optionally, the images acquired by the camera array and the corresponding annotated images are used to form training data according to preset classifications, and the step of training the target re-identification model using the training data includes:
[0010] Using images acquired by each camera in the camera array and corresponding annotated images to form a first training data set, and using the first training data set to train a convolutional neural network to obtain a target detection sub-model;
[0011] The second training data set is composed of images exposed in the camera array according to the clockwise sequence and the corresponding labeled images, and the third training data set is composed of images exposed in the camera array according to the counterclockwise sequence and the corresponding labeled images. The twin network is trained by using the second training data set and the third training data set to obtain a similarity calculation sub-model.
[0012] Optionally, the target detection submodel is cascaded with the similarity calculation submodel.
[0013] Optionally, the convolutional neural network is constructed using a YOLOv8 algorithm, wherein the weights of the Yolov8 algorithm are associated with exposure parameters of the camera array.
[0014] Optionally, the similarity calculation sub-model performs similarity calculation on two frames of images that are exposed sequentially in the exposure order based on targets contained in the images to obtain an image sequence arranged in descending order of similarity.
[0015] According to a second aspect of the embodiments of this specification, a surround view target re-identification system is provided, comprising a camera array and a recognition and tracking module, wherein
[0016] The camera array is configured to include a preset number of cameras arranged in a preset shape, wherein the preset number of cameras are evenly arranged along the circumference of a regular polygon to form a camera array, and the camera array exposes the cameras in sequence in a set exposure sequence, wherein the exposure sequence includes a clockwise sequence and a counterclockwise sequence, wherein one of any two adjacent cameras is designated as a clockwise starting camera and the other camera is designated as a counterclockwise starting camera, and exposure is sequentially performed with one camera interval in a clockwise direction starting from the clockwise starting camera to form a clockwise sequence, and exposure is sequentially performed with one camera interval in a counterclockwise direction starting from the counterclockwise starting camera to form a counterclockwise sequence, and the clockwise sequence and the counterclockwise sequence start exposure simultaneously or start exposure with a preset time difference;
[0017] The recognition and tracking module is configured to input the real-time image obtained from the preset camera array into a pre-trained target re-recognition model, form a target database with the targets recognized by the target re-recognition model, detect the input image to recognize the target in the image, and then determine the image sequence associated with the target according to the recognized target; obtain an initial image containing the target to be retrieved, input the initial image containing the target to be retrieved into the target re-recognition model to obtain the target to be retrieved, use the target re-recognition model to calculate the similarity of the target to be retrieved in the initial image in the target database, determine the associated targets of the target to be retrieved in the target database, and arrange the associated targets in descending order of similarity; according to the obtained associated targets, call out the image sequence containing the associated targets captured by the camera array according to the preset exposure sequence; obtain the movement trajectory of the target to be retrieved according to the timestamp of the image in the image sequence and the shooting position of each camera in the corresponding exposure sequence, so as to realize the recognition and tracking of the target to be retrieved.
[0018] According to a third aspect of an embodiment of this specification, there is provided a monitoring device, including a monitoring device and a computing device, wherein the monitoring device is used to acquire images within a monitoring range, and the computing device is used to process the images acquired by the monitoring device, the monitoring device includes a camera array in the surround view target re-identification system, and the computing device runs the steps of the surround view target re-identification method.
[0019] The beneficial effects of the embodiments of this specification are as follows:
[0020] The embodiments of the present specification provide a surround target re-identification method, system and monitoring equipment. Images are acquired by a camera array arranged in a ring in accordance with a preset exposure sequence, so as to obtain an image sequence that is correlated in both time and space sequences. The images are analyzed and processed by a pre-trained target re-identification model, and the identified targets are formed into a target database. When target tracking is required, the initial image to be retrieved is input into the target re-identification model to obtain the target to be retrieved. The target re-identification model is used to calculate the similarity of the target to be retrieved in the initial image in the target database, and the associated targets of the target to be retrieved in the target database are determined, and then the associated targets are arranged in order of similarity from high to low, and then the image sequence containing the associated targets captured by the camera array in the preset exposure sequence is called out, so as to obtain the movement trajectory of the target to be retrieved according to the timestamp of the image in the image sequence and the shooting position of each camera in the corresponding exposure sequence, so as to realize the recognition and tracking of the target to be retrieved.
[0021] The innovative features of the embodiments of this specification include:
[0022] 1. In this specification, the camera array adopts a regular polygon surround camera array, which has a 360-degree visual feature and can complete the field of view of the surround image without installing a motor to rotate. It has high reliability and is one of the innovative points of the embodiments of this specification.
[0023] 2. In this specification, the camera exposure time and sequence of the camera array adopts one group counterclockwise and one group clockwise, and the triggering method visually tracks the trajectory, which is convenient for tracking the target, which is one of the innovative points of the embodiments of this specification.
[0024] 3. In this specification, based on the arrangement of the camera array and the exposure sequence, an image sequence with spatial and temporal sequence associations is obtained, and then the images are used to train a target re-identification model. The target re-identification model uses a cascade of a detection convolutional neural network and a twin network, which is one of the innovative points of the embodiments of this specification. BRIEF DESCRIPTION OF THE DRAWINGS
[0025] In order to more clearly illustrate the technical solutions in the embodiments of this specification or related technologies, the drawings required for use in the embodiments or related technical descriptions will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0026] Figure 1 A flowchart of a surround target re-identification method provided in an embodiment of this specification;
[0027] Figure 2 A schematic diagram of the exposure sequence of a camera array provided in one embodiment of this specification;
[0028] Figure 3 A schematic diagram of the operation flow of the target re-identification model provided in one embodiment of this specification;
[0029] Figure 4 A schematic diagram of the structure of a surround target re-identification system provided in one embodiment of this specification;
[0030] Figure 5 A schematic diagram of the structure of a monitoring device provided in one embodiment of this specification. DETAILED DESCRIPTION
[0031] The following will be combined with the drawings in the embodiments of this specification to clearly and completely describe the technical solutions in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0032] It should be noted that the terms "including" and "having" and any variations thereof in the embodiments of this specification and the accompanying drawings are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device including a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units that are not listed, or may optionally include other steps or units that are inherent to these processes, methods, products or devices.
[0033] The embodiments of this specification disclose a surround target re-identification method, system and monitoring device provided by the embodiments of this specification, which are described in detail below.
[0034] Figure 1 The following is a flow chart of a method for re-identifying a surround-view target provided in one embodiment of the present specification. Figure 1 As shown, a surround target re-identification method includes:
[0035] S110, inputting a pre-trained target re-recognition model into a real-time image obtained from a preset camera array, and forming a target database with targets recognized by the target re-recognition model. The target re-recognition model detects the input image to recognize the target in the image, and then determines an image sequence associated with the target based on the recognized target.
[0036] In a specific embodiment, the camera array includes a preset number of cameras arranged in a preset shape, wherein
[0037] A preset number of cameras are evenly arranged along the circumference of the regular polygon to form a camera array, and the camera array exposes the cameras in sequence according to a set exposure sequence.
[0038] The exposure sequence includes a clockwise sequence and a counterclockwise sequence. One of any two adjacent cameras is designated as a clockwise starting camera and the other camera is designated as a counterclockwise starting camera. The clockwise starting camera is used as the starting point to sequentially expose the camera in a clockwise direction with one camera interval at a time to form a clockwise sequence. The counterclockwise starting camera is used as the starting point to sequentially expose the camera in a counterclockwise direction with one camera interval at a time to form a counterclockwise sequence. The clockwise sequence and the counterclockwise sequence start to expose at the same time or start to expose with a preset time difference. The preset number is an even number.
[0039] Figure 2 The following is a schematic diagram of the exposure sequence of a camera array provided in one embodiment of this specification. Figure 2 As shown, in one embodiment, a 16-eye camera array is used, which has a 360-degree visual feature and can complete the field of view of the surround image without installing a motor rotation, thereby having high reliability. According to the method of this specification, the camera exposure time and sequence adopts a counterclockwise group (letters a to h) and a clockwise group (numbers 1 to 8), and the triggering method visually tracks the trajectory, which is convenient for tracking the target.
[0040] In a specific implementation, before the step of inputting the real-time image obtained from the preset camera array into the pre-trained target re-identification model, the method further includes:
[0041] S102: Annotate images acquired from the camera array within a preset time length.
[0042] In specific implementation, the step of marking the images acquired from the camera array within a preset time period in S102 includes three methods, wherein:
[0043] The first method is to add a label frame to the image through manual cropping;
[0044] The second method is to process the image through an algorithm model with target detection function and add a label box to the image;
[0045] The third method is to process the image through an algorithm model with target tracking function and add a label box to the image.
[0046] S104, using the images acquired by the camera array and the corresponding annotated images to form training data according to preset classifications, and training and obtaining a target re-identification model through the training data, wherein the preset classifications are related to the arrangement shape of the camera array and the set exposure sequence.
[0047] In a specific embodiment, the step of S104, using the images acquired by the camera array and the corresponding annotated images to form training data according to preset classifications, and training the target re-identification model using the training data, includes:
[0048] S1041. Use the images acquired by each camera in the camera array and the corresponding labeled images to form a first training data set, and use the first training data set to train a convolutional neural network to obtain a target detection sub-model.
[0049] The convolutional neural network is constructed using the YOLOv8 algorithm, wherein the weights of the Yolov8 algorithm are associated with the exposure parameters of the camera array.
[0050] S1042. Use the images exposed in the camera array according to the clockwise sequence and the corresponding labeled images to form a second training data set, use the images exposed in the camera array according to the counterclockwise sequence and the corresponding labeled images to form a third training data set, train the twin network using the second training data set and the third training data set to obtain a similarity calculation sub-model.
[0051] The similarity calculation sub-model performs similarity calculation on two frames of images that are exposed sequentially in the exposure order based on the objects contained in the images, and obtains an image sequence arranged in descending order according to the similarity.
[0052] The target detection sub-model is cascaded with the similarity calculation sub-model.
[0053] Figure 3 The following is a schematic diagram of the operation flow of the target re-identification model provided in one embodiment of this specification. Figure 3 As shown in Figure 2, the algorithm flow of the target re-identification model is as follows: Figure 3First, collect data and generate original collected data based on 16 surround-view cameras. Secondly, generate bounding boxes and annotated training data, specifically including three ways to obtain bounding boxes (annotated boxes): 1. Manual cropping and annotation; 2. Target detection and annotation; 3. Tracking target annotation, to annotate the images obtained by 16 cameras. Then train the model to train a Re-ID model based on the annotated training data. This step is the core of the development of the Re-ID system. The last step is to retrieve the target. For the input retrieval object, use the Re-ID model obtained in the previous stage of training to extract the feature representation, calculate the similarity with the gallery and sort it from high to low, and you can get the retrieval image list. Going further, target re-identification can be divided into two stages: target detection + target image retrieval. After the target re-identification model is built, the algorithm flow of each step corresponds to the method flow of this manual as shown below. Figure 3 Marked as shown.
[0054] The target detection algorithm uses the YOLOv8 algorithm, and the target image retrieval algorithm uses the detected target results to build a large image library, and then uses the predicted image (probe) to match the corresponding target from the image library. The matching algorithm uses the siamese network to extract feature values to calculate the similarity and identify targets with the same identity.
[0055] S120, obtaining an initial image containing the target to be retrieved, inputting the initial image containing the target to be retrieved into a target re-identification model to obtain the target to be retrieved, using the target re-identification model to calculate the similarity of the target to be retrieved in the initial image in a target database, determining the associated targets of the target to be retrieved in the target database, and arranging the associated targets in descending order of similarity.
[0056] S130 . According to the obtained associated target, retrieve an image sequence including the associated target captured by the camera array according to a preset exposure sequence.
[0057] S140, obtaining a movement trajectory of the target to be retrieved according to the time stamps of the images in the image sequence and the shooting positions of each camera in the corresponding exposure sequence, so as to realize identification and tracking of the target to be retrieved.
[0058] Figure 4 This is a schematic diagram of the structure of a surround target re-identification system provided in one embodiment of this specification. Figure 4 As shown, a surround target re-identification system 400 includes a camera array 410 and a recognition and tracking module 420, wherein
[0059] The camera array 410 is configured to include a preset number of cameras arranged in a preset shape, wherein the preset number of cameras are evenly arranged along the circumference of a regular polygon to form a camera array, and the camera array exposes the cameras in sequence according to a set exposure sequence, and the exposure sequence includes a clockwise sequence and a counterclockwise sequence, and one of any two adjacent cameras is designated as a clockwise starting camera, and the other camera is designated as a counterclockwise starting camera. Starting from the clockwise starting camera, exposure is sequentially performed with one camera at a time in the clockwise direction to form a clockwise sequence, and starting from the counterclockwise starting camera, exposure is sequentially performed with one camera at a time in the counterclockwise direction to form a counterclockwise sequence. The clockwise sequence and the counterclockwise sequence start exposure at the same time or start exposure with a preset time difference.
[0060] The recognition and tracking module 420 is configured to input the real-time image obtained from the preset camera array into a pre-trained target re-recognition model, form a target database with the targets recognized by the target re-recognition model, detect the input image to recognize the target in the image, and then determine the image sequence associated with the target according to the recognized target; obtain an initial image containing the target to be retrieved, input the initial image containing the target to be retrieved into the target re-recognition model to obtain the target to be retrieved, use the target re-recognition model to calculate the similarity of the target to be retrieved in the initial image in the target database, determine the associated targets of the target to be retrieved in the target database, and arrange the associated targets in descending order of similarity; according to the obtained associated targets, call out the image sequence containing the associated targets captured by the camera array according to the preset exposure sequence; obtain the movement trajectory of the target to be retrieved according to the timestamp of the image in the image sequence and the shooting position of each camera in the corresponding exposure sequence, so as to realize the recognition and tracking of the target to be retrieved.
[0061] Figure 5 This is a schematic diagram of the structure of a monitoring device provided in an embodiment of this specification. Figure 3 As shown, a monitoring device 500 includes a monitoring device 510 and a computing device 520, wherein the monitoring device 510 is used to obtain images within a monitoring range, and the computing device 520 is used to process the images obtained by the monitoring device 510, and the monitoring device 510 includes a camera array in the surround view target re-identification system, and the computing device 520 runs the steps of the surround view target re-identification method.
[0062] To summarize, the embodiments of this specification provide a surround target re-identification method, system and monitoring equipment provided by the embodiments of this specification, which utilize a camera array and a re-identification model algorithm, adopt a cascade of a detection convolutional neural network and a twin network, and use a ring camera array to solve the recognition and retrieval of targets across cameras and scenes. It can recognize the target based on relevant information such as the target's appearance and posture, and has the potential to be widely used in scenarios such as smart city video surveillance, shopping malls, and criminal investigation video target retrieval.
[0063] Those skilled in the art can understand that the accompanying drawings are only schematic diagrams of an embodiment, and the modules or processes in the accompanying drawings are not necessarily required to implement the present invention.
[0064] Those skilled in the art can understand that the modules in the device in the embodiment can be distributed in the device in the embodiment according to the description of the embodiment, or can be changed accordingly and located in one or more devices different from the embodiment. The modules in the above embodiment can be combined into one module, or can be further divided into multiple sub-modules.
[0065] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A surround target re-identification method, characterized in that: include: Inputting the real-time images obtained from the preset camera array into a pre-trained target re-recognition model, the targets identified by the target re-recognition model form a target database, the target re-recognition model detects the input image to identify the target in the image, and then determines the image sequence associated with the target according to the identified target; Acquire an initial image containing a target to be retrieved, input the initial image containing the target to be retrieved into a target re-identification model to obtain the target to be retrieved, calculate the similarity of the target to be retrieved in the initial image in a target database using the target re-identification model, determine the associated targets of the target to be retrieved in the target database, and arrange the associated targets in descending order of similarity; According to the obtained associated target, calling out an image sequence including the associated target captured by the camera array according to a preset exposure sequence; Obtaining the movement trajectory of the target to be retrieved according to the time stamps of the images in the image sequence and the shooting positions of each camera in the corresponding exposure sequence, so as to realize the recognition and tracking of the target to be retrieved; The camera array includes a preset number of cameras arranged in a preset shape, wherein A preset number of cameras are evenly arranged along the circumference of a regular polygon to form a camera array, and the camera array exposes the cameras in sequence according to a set exposure sequence, and the exposure sequence includes a clockwise sequence and a counterclockwise sequence, and one of any two adjacent cameras is designated as a clockwise starting camera and the other camera is designated as a counterclockwise starting camera, and exposure is sequentially performed with one camera interval in a clockwise direction starting from the clockwise starting camera to form a clockwise sequence, and exposure is sequentially performed with one camera interval in a counterclockwise direction starting from the counterclockwise starting camera to form a counterclockwise sequence, and the clockwise sequence and the counterclockwise sequence start exposure at the same time or start exposure with a preset time difference; Before the step of inputting the real-time image obtained from the preset camera array into the pre-trained target re-identification model, the method further includes: Annotating images acquired from the camera array within a preset time length; The images acquired by the camera array and the corresponding annotated images are used to form training data according to preset classifications, and a target re-identification model is trained and obtained through the training data, wherein the preset classifications are related to the arrangement shape of the camera array and the set exposure sequence; The step of using the images acquired by the camera array and the corresponding annotated images to form training data according to preset classifications, and training and obtaining a target re-identification model through the training data includes: The first training data set is formed by using images acquired by each camera in the camera array and the corresponding annotated images, and a convolutional neural network is trained by using the first training data set to obtain a target detection sub-model; the convolutional neural network is constructed by using a YOLOv8 algorithm, wherein the weight of the Yolov8 algorithm is associated with the exposure parameter of the camera array; The second training data set is composed of images exposed in the camera array according to the clockwise sequence and the corresponding labeled images, and the third training data set is composed of images exposed in the camera array according to the counterclockwise sequence and the corresponding labeled images. The twin network is trained by using the second training data set and the third training data set to obtain a similarity calculation sub-model.
2. The method according to claim 1, characterized in that The step of marking the images acquired from the camera array within a preset time period includes three methods, wherein: The first method is to add a label frame to the image through manual cropping; The second method is to process the image through an algorithm model with target detection function and add a label box to the image; The third method is to process the image through an algorithm model with target tracking function and add a label box to the image.
3. The method according to claim 1, characterized in that The target detection sub-model is cascaded with the similarity calculation sub-model.
4. A surround target re-identification system, characterized in that: It includes a camera array and a recognition and tracking module, wherein The camera array is configured to include a preset number of cameras arranged in a preset shape, wherein the preset number of cameras are evenly arranged along the circumference of a regular polygon to form a camera array, and the camera array exposes the cameras in sequence according to a set exposure sequence, wherein the exposure sequence includes a clockwise sequence and a counterclockwise sequence, and one of any two adjacent cameras is designated as a clockwise starting camera and the other camera is designated as a counterclockwise starting camera, and exposure is sequentially performed in a clockwise direction with one camera interval at a time starting from the clockwise starting camera to form a clockwise sequence, and exposure is sequentially performed in a counterclockwise direction with one camera interval at a time starting from the counterclockwise starting camera to form a counterclockwise sequence, and the clockwise sequence and the counterclockwise sequence start to expose at the same time or start to expose with a preset time difference; the camera array includes a preset number of cameras arranged in a preset shape, wherein A preset number of cameras are evenly arranged along the circumference of a regular polygon to form a camera array, and the camera array exposes the cameras in sequence according to a set exposure sequence, and the exposure sequence includes a clockwise sequence and a counterclockwise sequence, and one of any two adjacent cameras is designated as a clockwise starting camera and the other camera is designated as a counterclockwise starting camera, and exposure is sequentially performed with one camera interval in a clockwise direction starting from the clockwise starting camera to form a clockwise sequence, and exposure is sequentially performed with one camera interval in a counterclockwise direction starting from the counterclockwise starting camera to form a counterclockwise sequence, and the clockwise sequence and the counterclockwise sequence start exposure at the same time or start exposure with a preset time difference; The recognition and tracking module is configured to input a pre-trained target re-recognition model into a real-time image obtained from a preset camera array, and form a target database with the targets recognized by the target re-recognition model. The target re-recognition model detects the input image to recognize the target in the image, and then determines an image sequence associated with the target according to the recognized target; Acquire an initial image containing a target to be retrieved, input the initial image containing the target to be retrieved into a target re-identification model to obtain the target to be retrieved, use the target re-identification model to calculate the similarity of the target to be retrieved in the initial image in a target database, determine the associated targets of the target to be retrieved in the target database, and arrange the associated targets in descending order of similarity; call out an image sequence containing the associated targets captured by the camera array according to a preset exposure sequence according to the obtained associated targets; obtain a movement trajectory of the target to be retrieved according to the timestamps of the images in the image sequence and the shooting positions of each camera in the corresponding exposure sequence, so as to realize the recognition and tracking of the target to be retrieved; Before the step of inputting the real-time image obtained from the preset camera array into the pre-trained target re-identification model, the method further includes: Annotating images acquired from the camera array within a preset time length; The images acquired by the camera array and the corresponding annotated images are used to form training data according to preset classifications, and a target re-identification model is trained and obtained through the training data, wherein the preset classifications are related to the arrangement shape of the camera array and the set exposure sequence; The step of using the images acquired by the camera array and the corresponding annotated images to form training data according to preset classifications, and training and obtaining a target re-identification model through the training data includes: The first training data set is formed by using images acquired by each camera in the camera array and the corresponding annotated images, and a convolutional neural network is trained by using the first training data set to obtain a target detection sub-model; the convolutional neural network is constructed by using a YOLOv8 algorithm, wherein the weight of the Yolov8 algorithm is associated with the exposure parameter of the camera array; The second training data set is composed of images exposed in the camera array according to the clockwise sequence and the corresponding labeled images, and the third training data set is composed of images exposed in the camera array according to the counterclockwise sequence and the corresponding labeled images. The twin network is trained by using the second training data set and the third training data set to obtain a similarity calculation sub-model.
5. A monitoring device, characterized in that: The invention comprises a monitoring device and a computing device, wherein the monitoring device is used to acquire images within a monitoring range, and the computing device is used to process the images acquired by the monitoring device. The monitoring device comprises a camera array as described in any one of claims 1 to 3, and the computing device runs the steps of a surround view target re-identification method as described in any one of claims 1 to 3.
