A multi-target detection and tracking system based on deep learning and a method thereof

Abstract

The application provides a kind of multi-target detection tracking system and method based on deep learning, including acquisition module: the module is used to AGV car is photographed, and image information is collected;Detection module: the target in the image information collected is detected using deep learning model, and the appearance feature of multiple AGV cars is extracted;Tracking module: using Kalman filtering algorithm, the motion trajectory of the last time trajectory of multiple AGV cars in new time is predicted, finally using four-stage matching algorithm is associated calculation, complete multi-target tracking;Output module: the system outputs and displays tracking result, from high confidence object matching to low confidence object matching, to historical position matching, and unconfirmed trajectory matching, each stage is optimized for specific situation, compared with the traditional only single high confidence object matching is optimized tracking, this not only can improve the accuracy of matching, also can maximumly reduce the situation of missing detection and false detection.

Description

Technical Field

[0001] This invention relates to the field of target tracking technology, and in particular to a multi-target detection and tracking system and method based on deep learning. Background Technology

[0002] In AGV (Automated Guided Vehicle) applications, deep learning-based multi-object detection and tracking systems can significantly enhance the autonomous navigation and transport capabilities of AGVs in complex environments. This system is a complex yet efficient computer vision system that combines deep learning and object detection and tracking algorithms to accurately detect and track multiple targets in consecutive video frames. The following outlines some core concepts and basic principles of this system. First, object detection is one of the key steps. By using deep learning object detection algorithms, the system can detect all targets of interest in each frame and obtain their positional information. These algorithms extract image features through convolutional neural networks, thereby identifying the position and shape of the target objects. Second, feature extraction is another crucial step. For each detected target, the system extracts its visual features (such as color, texture, and shape) or motion features (such as speed and acceleration). These features play a key role in subsequent tracking, helping to match and identify the same targets in consecutive frames.

[0003] In existing multi-target detection and tracking systems, especially in environments involving the interactive operation of multiple AGVs, mutual occlusion frequently occurs. This increases the difficulty of target detection, especially since the confidence score of occluded targets drops significantly. Consequently, when traditional tracking algorithms that rely solely on high-confidence bounding boxes are used again, these low-confidence targets are easily overlooked, leading to interruptions or even loss of the target trajectory, thus affecting the stability and accuracy of tracking.

[0004] Therefore, it is necessary to provide a new deep learning-based multi-target detection and tracking system and method to solve the above-mentioned technical problems. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a multi-target detection and tracking system and method based on deep learning.

[0006] The present invention provides a deep learning-based multi-target detection and tracking system and method, comprising: A deep learning-based multi-target detection and tracking system includes: Data Acquisition Module: This module is used to capture images of the AGV (Automated Guided Vehicle) and collect image information. Detection module: Uses deep learning models to detect targets in the acquired image information and extracts appearance features from multiple AGV vehicles; Tracking module: Collects the appearance features of the AGVs in the detection module, uses the Kalman filter algorithm to predict the motion trajectory of multiple AGVs from the previous moment to the new moment, and finally uses the processing unit to perform correlation calculation to complete multi-target tracking; Output module: The system collects the multi-target tracking results from the tracking module and displays the tracking results.

[0007] Furthermore, the data association method is the BYTE data association method, and its specific usage process is as follows: S01: First, each detection box is divided into two categories: high-scoring boxes and low-scoring boxes, and then matched twice; S02: The first matching uses the high-resolution bounding box and the previous tracking trajectory. The second matching uses the low-resolution bounding box and the tracking trajectory that did not match the high-resolution bounding box in the first matching.

[0008] Furthermore, the specific steps for extracting the appearance features of multiple AGV vehicles in the detection module are as follows: S001: Receives image information acquired by the acquisition module, annotates the acquired images, marks the position and bounding box of the AGV, and forms a dataset for training the YOLO model; S002: Train the selected YOLO model using the labeled dataset, and adjust the model parameters and optimizer settings to enable the model to accurately identify AGV vehicles; S003: Input the image to be detected into the trained model. The model will output the position and confidence of the AGV in the image. The overlapping bounding boxes are removed by the non-maximum suppression method to obtain the final detection result. S004: Based on object detection, a pre-trained convolutional neural network can be used as a feature extractor to extract the appearance features of the AGV vehicle.

[0009] Furthermore, the device used to acquire images in the acquisition module is a camera.

[0010] Furthermore, the tracking results output by the output module are in the form of images.

[0011] Furthermore, the processing unit performs correlation calculations to complete multi-target tracking in the following steps: S1: Divide the time period of the AGV's travel into the trajectory set at time t-1, and divide it into tracking trajectory, unconfirmed trajectory, and lost trajectory. The tracking trajectory and lost trajectory are predicted by Kalman filter to obtain the predicted position of the trajectory at time t. S2: High-confidence object matching, based on BYTE data association, will perform high-confidence object matching between the predicted position information of the AGV at time t and the high-resolution detected object; S3: For trajectories that fail to match high-confidence objects, match them with low-confidence detected objects. Trajectories that fail to match are matched with low-score detected objects to perform low-confidence object matching. S4: Historical position matching. For trajectories that fail to match twice in the first two times, the historical position of the detected object is matched. This historical position is the position of the AGV vehicle in time period t-1. S5: Unconfirmed trajectory matching. For high-confidence objects among the detected objects that have also failed to match historical locations, match them with unconfirmed trajectories. S6: Based on the results of the four successfully matched trajectories, the system updates the trajectory set of the AGV at time t. The successfully matched trajectories will be updated according to the latest position and status information; the unmatched trajectories will be marked as lost or removed from the current tracking set; at the same time, newly detected objects that have not yet been associated will be added to the unconfirmed trajectory set.

[0012] Furthermore, the matching method used in S2 and S3, which uses high-confidence object matching and low-confidence detection object matching, adopts the BIoU motion similarity metric. By emphasizing boundary matching, the BIoU motion similarity metric can more accurately capture the contour changes of moving targets, thereby improving the accuracy of moving target detection and tracking.

[0013] The deep learning-based multi-object detection method, applied to the aforementioned deep learning-based multi-object detection and tracking system, includes the following steps: S0001: First, the AGV vehicle is photographed by the acquisition module, and the image information of the AGV vehicle is acquired; S0002: Receives image information acquired by the acquisition module, annotates the acquired images, marks the position and bounding box of the AGV, forming a dataset for training the YOLO model. The selected YOLO model is trained using the labeled dataset. By adjusting the model parameters and optimizer settings, the model can accurately identify the AGV. The image to be detected is input into the trained model, and the model will output the position and confidence of the AGV in the image. Overlapping bounding boxes are removed using the non-maximum suppression method to obtain the final detection result. Based on object detection, a pre-trained convolutional neural network can be used as a feature extractor to extract the appearance features of the AGV. S0003: Subsequently, the tracking module receives the appearance feature data of the AGV vehicle, uses the Kalman filter algorithm to predict the motion trajectory of multiple AGV vehicles from the previous moment to the new moment, and finally uses the four-stage matching algorithm to perform association calculation to complete multi-target tracking. S0004: The tracking results are displayed through the output module in the form of images.

[0014] Compared with related technologies, the deep learning-based multi-target detection and tracking system and method provided by this invention have the following advantages: This invention uses the Kalman filter algorithm to predict the trajectory of multiple AGVs from the previous moment to the new moment. Finally, a four-stage matching algorithm is used for association calculation to complete multi-target tracking. In specific use, the time period of AGV travel is divided into the trajectory set at time t-1, which is divided into tracking trajectory, unconfirmed trajectory, and lost trajectory. For trajectories that fail to match high-confidence objects, they are matched with low-confidence detection objects. Specifically, the failed matching trajectory is matched with low-confidence detection objects. For trajectories that fail to match twice, they are matched with the detection objects based on their historical positions. For high-confidence objects among the detection objects that also fail to match their historical positions, they are matched with unconfirmed trajectories. Based on the results of the above four successfully matched trajectories, the system updates the trajectory set of the AGV at time t. The successfully matched trajectories will be updated according to the latest position and status information. The four-stage matching algorithm can comprehensively and meticulously handle trajectory matching problems under different conditions, from high-confidence object matching to low-confidence object matching, then to historical location matching, and unconfirmed trajectory matching. Each stage is optimized for specific situations. Compared with the traditional method of only optimizing the tracking of high-confidence object matching, this can not only improve the accuracy of matching, but also minimize the occurrence of missed detections and false detections. Attached Figure Description

[0015] Figure 1 This is an overall block diagram of the deep learning-based multi-target detection and tracking system provided by the present invention; Figure 2 A flowchart of the four-stage matching algorithm for the deep learning-based multi-target detection and tracking system provided by the present invention; Figure 3 The flowchart of the deep learning-based multi-target detection and tracking method provided by this invention is shown. Detailed Implementation

[0016] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0017] Please refer to the following: Figure 1 , Figure 2, Figure 3 ,in, Figure 1 This is an overall block diagram of the deep learning-based multi-target detection and tracking system provided by the present invention; Figure 2 A flowchart of the four-stage matching algorithm for the deep learning-based multi-target detection and tracking system provided by the present invention; Figure 3 The flowchart of the deep learning-based multi-target detection and tracking method provided by this invention is shown.

[0018] In the specific implementation process, such as Figures 1-3 As shown, a deep learning-based multi-target detection and tracking system includes: Data Acquisition Module: This module is used to capture images of the AGV (Automated Guided Vehicle) and collect image information. Detection module: Uses deep learning models to detect targets in the acquired image information and extracts appearance features from multiple AGV vehicles; Tracking module: Collects the appearance features of the AGVs in the detection module, uses the Kalman filter algorithm to predict the motion trajectory of multiple AGVs from the previous moment to the new moment, and finally uses the processing unit to perform correlation calculation to complete multi-target tracking; Output module: The system collects the multi-target tracking results from the tracking module and displays the tracking results. The processing unit performs association calculations, and the multi-target tracking method is a four-stage matching algorithm, with the specific steps as follows: S1: Divide the time period of the AGV's travel into the trajectory set at time t-1, and divide it into tracking trajectory, unconfirmed trajectory, and lost trajectory. The tracking trajectory and lost trajectory are predicted by Kalman filter to obtain the predicted position of the trajectory at time t. S2: High-confidence object matching, based on BYTE data association, will perform high-confidence object matching between the predicted position information of the AGV at time t and the high-resolution detected object; S3: For trajectories that fail to match high-confidence objects, match them with low-confidence detected objects. Trajectories that fail to match are matched with low-score detected objects to perform low-confidence object matching. S4: Historical position matching. For trajectories that fail to match twice in the first two times, the historical position of the detected object is matched. This historical position is the position of the AGV vehicle in time period t. In the specific implementation process, if the first two matching attempts fail, the system will try to match using historical location information. Historical location is the known location of the AGV within the time period t. These locations may come from previous sensor data or trajectory records. By comparing the historical location with the current location of the detected object, the system can further narrow down the matching range and improve the accuracy of the matching. S5: Unconfirmed trajectory matching. For high-confidence objects among the detected objects that have also failed to match historical locations, match them with unconfirmed trajectories. In the specific implementation process, for those high-confidence detection objects that cannot find a match in the historical location matching, the system matches them with unconfirmed trajectories. Unconfirmed trajectories are those newly detected trajectories that have not yet been matched with any known trajectories. By matching with unconfirmed trajectories, the system may be able to discover new AGV vehicles that have not been noticed before. S6: Based on the results of the above four matching stages, the system updates the trajectory set at time t. Successfully matched trajectories will be updated according to the latest position and status information; unsuccessfully matched trajectories may be marked as lost or removed from the current tracking set; at the same time, newly detected, yet unassociated objects will be added to the unconfirmed trajectory set. The specific process of the Kalman filter algorithm mentioned above is as follows, where the state vector D of the Kalman filter is: ,in, and These are the center coordinates of the target bounding box; It is the aspect ratio of the target bounding box; The length of the target bounding box; They represent the corresponding parameters respectively. speed, The target's lateral and longitudinal accelerations; Let X = Y= Z= The target information of the AGV is then represented as follows:

[0019] The prediction process of the Kalman filter algorithm mentioned above is as follows:

[0020] in, This is the predicted value of the AGV's position information in the previous frame. Let Z be the optimal estimate of the AGV's position information, B be the AGV's state transition matrix, P be the AGV's estimated covariance matrix, and Q be the AGV's state transition covariance matrix. The state transition matrix, as mentioned above, describes how the system state transitions from the previous time step to the current time step. The estimated covariance matrix is ​​used to quantify the uncertainty of the AGV's state estimation; this matrix contains information about the covariance between the AGV's state variables, describing the accuracy of the state estimation and the correlation between the state variables. The state transition covariance matrix describes the uncertainty of the AGV's predicted state; this matrix contains information about the covariance between the AGV's state variables and is used to estimate the uncertainty of the state variables and their correlation. The Kalman filter algorithm predicts the target position information of the current frame based on the target information of the previous frame, and compares it with the target position information output by the detection stage. The application of the Kalman filter algorithm can significantly improve the accuracy and stability of target tracking. It can predict the motion state of the AGV at the new moment based on the trajectory information of the AGV at the previous moment, effectively handle interference factors such as occlusion and noise that may occur during the target movement, reduce the loss of the target, and further improve the tracking accuracy. The specific usage process of the BYTE data association method is as follows: S01: First, divide each detection box into two categories: high-scoring boxes and low-scoring boxes, and perform two matching operations; S02: The first matching is done by using the high-resolution bounding box and the previous tracking trajectory. The second matching is done by using the low-resolution bounding box and the tracking trajectory that did not match the high-resolution bounding box in the first matching. The above process is particularly helpful in situations where objects are occluded, by detecting the overlap between the bounding box and the trajectory, and thus extracting the real object from the low-scoring box, thereby maintaining the continuity of the trajectory. The specific steps for extracting the appearance features of multiple AGV vehicles in the detection module are as follows: S001: Receives image information acquired by the acquisition module, annotates the acquired images, marks the position and bounding box of the AGV, and forms a dataset for training the YOLO model; S002: Train the selected YOLO model using the labeled dataset, and adjust the model parameters and optimizer settings to enable the model to accurately identify AGV vehicles; S003: Input the image to be detected into the trained model. The model will output the position and confidence of the AGV in the image. The overlapping bounding boxes are removed by the non-maximum suppression method to obtain the final detection result. S004: Based on object detection, a pre-trained convolutional neural network can be used as a feature extractor to extract the appearance features of the AGV vehicle.

[0021] The device for acquiring images in the acquisition module is a camera, and the tracking results output in the output module are in the form of images. The matching method used in S2 and S3, which uses high-confidence object matching and low-confidence detection object matching, adopts the BIoU motion similarity metric. The BIoU motion similarity measurement method is as follows: First, the boundary lines of the predicted bounding box and the ground truth bounding box are determined, and the intersection length between these two boundary lines is calculated. This can be achieved by finding the overlapping part of the two boundary lines and calculating its length. Then, the total length of the boundary lines of the predicted bounding box and the ground truth bounding box is calculated by adding the lengths of the two boundary lines. Finally, the boundary intersection length is divided by the boundary union length to obtain the BIoU value. The higher the BIoU value, the higher the degree of boundary matching between the predicted bounding box and the ground truth bounding box, i.e., the higher the similarity. Through the BIoU motion similarity measurement method, since BIoU pays special attention to the degree of boundary matching between the predicted bounding box and the ground truth bounding box, in the AGV motion scene, the boundary of the AGV often carries important information, such as the shape and posture of the AGV. Therefore, by emphasizing boundary matching, BIoU can more accurately capture the contour changes of moving targets, thereby improving the accuracy of moving target detection and tracking. The BIoU motion similarity metric described above is traditionally measured using IoU, which is calculated as IoU = Intersection Area / (Predicted Area + True Area - Intersection Area). Here, the intersection area is the area of ​​the intersection between the predicted and true results, the predicted area is the area of ​​the predicted result, and the true area is the area of ​​the true result. By adopting the BIoU motion similarity metric, BIoU not only considers the overlap between target bounding boxes but also incorporates the motion information of the AGV vehicle. This allows BIoU to more comprehensively evaluate the similarity of targets across consecutive frames. Especially when the AGV vehicle moves quickly or there are multiple AGV vehicle targets in the scene, BIoU can provide more accurate measurement results, helping to reduce AGV vehicle loss and false detection during tracking, and further improving the performance of the entire tracking system. It should be noted that the specific training method for the above YOLO model is as follows: First, a training dataset needs to be prepared. This dataset should contain a large number of labeled images, with each AGV (Automated Guided Vehicle) accurately labeled. Typically, the dataset is divided into training, validation, and test sets to validate the model's performance during training and ultimately evaluate its accuracy. Next, install the necessary deep learning frameworks and dependencies, selecting an appropriate YOLO version (YOLOv7 is chosen here). Then, configure the model's parameters, such as learning rate, batch size, and number of iterations. Train the model using the training set. During training, the model learns how to recognize target objects in images and optimizes its internal parameters to improve recognition accuracy. During training, use the validation set to evaluate the model's performance. Based on the validation results, adjust the model's parameters or structure to optimize its performance. This process may require multiple iterations until the model's performance on the validation set is satisfactory. Finally, use the test set, independent of the training and validation sets, to test and evaluate the trained model. The test set is used to ultimately assess the model's generalization ability and accuracy. In the specific implementation process, the convolutional neural network process used above is as follows: The image is first input for preprocessing to make it into a standard size; the preprocessed image is input into the encoder to extract features and generate a high-resolution feature map; the high-resolution feature map is passed through the object detection branch network and the appearance feature extraction branch network to generate the appearance features of the AGV vehicle. The deep learning-based multi-object detection method, applied to the aforementioned deep learning-based multi-object detection and tracking system, includes the following steps: S0001: First, the AGV vehicle is photographed by the acquisition module, and the image information of the AGV vehicle is acquired; S0002: Receives image information acquired by the acquisition module, annotates the acquired images, marks the position and bounding box of the AGV, forming a dataset for training the YOLO model. The selected YOLO model is trained using the labeled dataset. By adjusting the model parameters and optimizer settings, the model can accurately identify the AGV. The image to be detected is input into the trained model, and the model will output the position and confidence of the AGV in the image. Overlapping bounding boxes are removed using the non-maximum suppression method to obtain the final detection result. Based on object detection, a pre-trained convolutional neural network can be used as a feature extractor to extract the appearance features of the AGV. S0003: Subsequently, the tracking module receives the appearance feature data of the AGV vehicle, uses the Kalman filter algorithm to predict the motion trajectory of multiple AGV vehicles from the previous moment to the new moment, and finally uses the four-stage matching algorithm to perform association calculation to complete multi-target tracking. S0004: The tracking results are displayed through the output module, and the tracking results are presented in the form of images; The method of removing overlapping bounding boxes using non-maximum suppression in step S0002 above is implemented as follows: A1: First, sort the bounding boxes of all detected AGVs according to their confidence scores (from highest to lowest) to ensure that bounding boxes with higher confidence scores are processed first in subsequent processing.

[0022] A2: Select the bounding box with the highest confidence from the sorted list and use it as the current bounding box to be processed; A3: Calculate the intersection-union ratio (IoU) between the currently selected bounding box and the remaining bounding boxes in the list. IoU is a metric that measures the degree of overlap between two bounding boxes, ranging from 0 to 1. The higher the IoU, the greater the degree of overlap between the two bounding boxes.

[0023] A4: If the IoU value between the currently selected bounding box and other bounding boxes in the list is greater than a preset threshold, then these overlapping bounding boxes will be removed from the list. This means that if two bounding boxes overlap too much, only the one with the highest confidence will be kept. A5: Continue to select the bounding box with the second highest confidence from the remaining bounding boxes, and repeat steps 3 and 4 until all bounding boxes have been traversed; The above-mentioned non-maximum suppression method ensures that each target is typically detected only once, which greatly reduces redundant detection boxes and improves the accuracy and efficiency of AGV detection. It also ensures that in the post-processing stage of AGV detection, only the relatively best bounding box is output as the final detection result, thereby reducing the possibility of false detection and duplicate detection.

[0024] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented using software plus a general-purpose hardware platform, or of course, using hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0025] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or basic characteristics. Therefore, the embodiments should be considered exemplary and non-limiting in all respects. The scope of the invention is defined by the appended claims rather than the foregoing description. Therefore, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.

[0026] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A multi-target detection and tracking system based on deep learning, characterized in that, include: Data Acquisition Module: This module is used to capture images of the AGV (Automated Guided Vehicle) and collect image information. Detection module: Uses deep learning models to detect targets in the acquired image information and extracts appearance features from multiple AGV vehicles; Tracking module: Collects the appearance features of the AGVs in the detection module, uses the Kalman filter algorithm to predict the motion trajectory of multiple AGVs from the previous moment to the new moment, and finally uses the processing unit to perform correlation calculation to complete multi-target tracking; Output module: The system collects the multi-target tracking results from the tracking module and displays the tracking results. The processing unit performs association calculations to complete multi-target tracking in the following steps: S1: Divide the trajectory set at time t-1 during the AGV's travel time into tracking trajectory, unconfirmed trajectory, and lost trajectory, and use a prediction algorithm to obtain the predicted position of the trajectory at time t; S2: High-confidence object matching, which uses data association to match the predicted position information of the AGV at time t with the high-resolution detected object. S3: For trajectories that fail to match high-confidence objects, match them with low-confidence detected objects. Trajectories that fail to match are matched with low-score detected objects to perform low-confidence object matching. S4: Historical position matching. For trajectories that fail to match in the first two attempts, match them with the historical position of the detected object. This historical position is the position of the AGV at time t-1. S5: Unconfirmed trajectory matching. For high-confidence objects among the detected objects that have also failed to match historical locations, match them with unconfirmed trajectories. S6: Based on the results of the successfully matched trajectories in steps S2-S5, the system updates the AGV's trajectory set at time t. The successfully matched trajectories will be updated based on the latest position and status information.

2. The deep learning-based multi-target detection and tracking system according to claim 1, characterized in that, The specific steps for extracting the appearance features of multiple AGV vehicles in the detection module are as follows: S001: Receive image information acquired by the acquisition module, annotate the acquired images, mark the position and bounding box of the AGV, form a dataset, use the annotated dataset to train the selected model, and adjust the model parameters and optimizer settings; S002: Input the image to be detected into the trained model, remove overlapping bounding boxes using the non-maximum suppression method, and obtain the final detection result; S003: Based on target detection, a feature extractor is used to extract appearance features.

3. The deep learning-based multi-target detection and tracking system according to claim 2, characterized in that, The device used to acquire images in the acquisition module is a camera. 4.The deep learning based multi-target detection and tracking system of claim 3, wherein, The tracking results output by the output module are in image format.

5. The deep learning-based multi-target detection and tracking system according to claim 4, characterized in that, In S2 and S3, the matching method that uses high-confidence object matching and low-confidence detection object matching adopts a motion similarity measurement algorithm.

6. A deep learning-based multi-target detection method, applied to any one of the deep learning-based multi-target detection and tracking systems of claims 1-5, characterized in that, Includes the following steps: S0001: First, the AGV vehicle is photographed by the acquisition module, and the image information of the AGV vehicle is acquired; S0002: Receives image information acquired by the acquisition module, annotates the acquired images, marks the position and bounding box of the AGV, forms a dataset, trains the selected model using the labeled dataset, and enables the model to accurately identify the AGV by adjusting the model parameters and optimizer settings. The image to be detected is input into the trained model, and the model will output the position and confidence of the AGV in the image. Overlapping bounding boxes are removed by non-maximum suppression method to obtain the final detection result. S0003: Subsequently, the tracking module receives the appearance feature data of the AGV vehicles, uses the Kalman filter algorithm to predict the motion trajectory of multiple AGV vehicles from the previous moment to the new moment, and finally uses a four-stage matching algorithm to perform association calculations to complete multi-target tracking. The tracking results are presented in image form and output through the output module. The four-stage matching algorithm is as follows: S1: Divide the trajectory set at time t-1 during the AGV's travel time into tracking trajectory, unconfirmed trajectory, and lost trajectory, and use a prediction algorithm to obtain the predicted position of the trajectory at time t; S2: High-confidence object matching, which uses data association to match the predicted position information of the AGV at time t with the high-resolution detected object. S3: For trajectories that fail to match high-confidence objects, match them with low-confidence detected objects. Trajectories that fail to match are matched with low-score detected objects to perform low-confidence object matching. S4: Historical position matching. For trajectories that fail to match in the first two attempts, match them with the historical position of the detected object. This historical position is the position of the AGV at time t-1. S5: Unconfirmed trajectory matching. For high-confidence objects among the detected objects that have also failed to match historical locations, match them with unconfirmed trajectories. S6: Based on the results of the successfully matched trajectories in steps S2-S5, the system updates the AGV's trajectory set at time t. The successfully matched trajectories will be updated based on the latest position and status information.

7. The method for multi-target detection based on deep learning according to claim 6, characterized in that, In S0002, overlapping bounding boxes are removed using the non-maximum suppression method to obtain the final detection result. Based on the target detection, a feature extractor is used to extract the appearance features of the AGV vehicle.

Images