CN117636267A - Vehicle target tracking method, system, device and storage medium - Google Patents

Abstract

The invention discloses a vehicle target tracking method, a vehicle target tracking system, a vehicle target tracking device and a storage medium, and relates to the technical field of artificial intelligence. And sequentially inputting each frame of image of the monitoring video into a YOLOv7 detector to perform target detection to obtain vehicle detection frame information, wherein a convolution block attention module is arranged in an ELAN structure of a main network layer of the YOLOv7 detector, and performs global maximum pooling and global average pooling operation on the input feature images by the convolution block attention module to obtain an output feature image fusing space-time attention features, and the feature representation capability is improved by the convolution block attention module in the ELAN structure, so that the vehicle target detection accuracy of the detector is improved. Inputting the information of the vehicle detection frame into a deep Sort tracker for track tracking to obtain a vehicle target tracking track, correcting the vehicle position coordinates in the vehicle target tracking track according to camera parameters to obtain a corrected vehicle target tracking track, and improving the accuracy of vehicle tracking.

Description

Vehicle target tracking method, system, device and storage medium

Technical field

The present invention relates to the field of artificial intelligence technology, and in particular to a vehicle target tracking method, system, device and storage medium.

Background technique

With the rapid development of smart transportation, the detection and tracking of vehicle targets through computer vision technology is increasingly important. This technology can provide information about the location, trajectory, speed and other information of the vehicle. In related technologies, target tracking solutions mainly adopt solutions based on joint detection and tracking. The detection-based tracking solution is to first perform target detection on each frame of the video sequence to obtain all targets in the image, and then convert it into a target association problem between the two frames before and after, construct a similarity matrix through IoU, appearance, etc., and Solve through Hungarian algorithm, greedy algorithm, etc. However, the target detection model used in the vehicle target tracking process has low accuracy and efficiency, which affects trajectory recognition. Moreover, because the target information reflected in the collected images is different from the real target, the accuracy of trajectory recognition is low.

Contents of the invention

The present invention aims to solve at least one of the technical problems existing in the prior art. To this end, the present invention proposes a vehicle target tracking method, system, device and storage medium, which can improve the accuracy of vehicle trajectory identification.

On the one hand, embodiments of the present invention provide a vehicle target tracking method, including the following steps:

Obtain surveillance video;

Each frame of the surveillance video is input into the YOLOv7 detector in turn for target detection to obtain vehicle detection frame information. The ELAN structure of the backbone network layer of the YOLOv7 detector is provided with a convolution block attention module. The block attention module performs global maximum pooling and global average pooling operations on the input feature map to obtain an output feature map that fuses spatiotemporal attention features;

Input the vehicle detection frame information into the DeepSort tracker for trajectory tracking to obtain the vehicle target tracking trajectory;

The vehicle position coordinates in the vehicle target tracking trajectory are corrected according to the camera parameters to obtain a corrected vehicle target tracking trajectory.

According to some embodiments of the present invention, the ELAN structure includes a first branch, a second branch, a third branch and a fourth branch, and both the first branch and the second branch are 1×1 convolution layers, so The third branch is a 1×1 convolution layer and a 3×3 convolution layer, and the fourth branch is a 1×1 convolution layer, a 3×3 convolution layer, and a 3×3 convolution layer. The convolutional layer; the first branch, the second branch, the third branch and the fourth branch are all connected to the splicing module.

According to some embodiments of the present invention, the convolution block attention module is provided between the first branch and the splicing module. The convolution block attention module includes a channel attention mechanism unit and a spatial attention unit connected in sequence. Force mechanism unit.

According to some embodiments of the present invention, the channel attention mechanism unit is used for:

The input features of the channel attention mechanism unit are pooled in the channel dimension respectively, and the channel global maximum pooling value and the channel global average pooling value are obtained;

The channel global maximum pooling value and the channel global average pooling value are respectively input into the fully connected neural network and then activated to obtain the first channel feature and the second channel feature;

The first channel feature and the second channel feature are added and activated to obtain the channel attention feature;

Perform a multiplication operation on the channel attention feature and the input feature of the input channel attention mechanism unit to obtain the output feature of the channel attention mechanism unit;

The spatial attention mechanism unit is used for:

The input features of the spatial attention mechanism unit are pooled in the spatial dimension respectively to obtain the spatial global maximum pooling features and the spatial global average pooling features;

After splicing the spatial global maximum pooling features and the spatial global average pooling features, perform dimensionality reduction and activation operations to obtain spatial attention features;

The spatial attention feature and the input feature of the spatial attention mechanism unit are multiplied to obtain the output feature of the spatial attention mechanism unit.

According to some embodiments of the present invention, the YOLOv7 detector is trained through a deep learning algorithm, and the model loss during the training process of the YOLOv7 detector is obtained through the following steps:

According to the degree of overlap between the predicted box output by the YOLOv7 detector and the real box, the intersection ratio loss is determined;

Determine the loss weight based on the distance measure between the predicted box and the real box output by the YOLOv7 detector;

Based on the attention mechanism, the model loss is determined based on the loss weight and the intersection-union ratio loss.

According to some embodiments of the present invention, inputting the vehicle detection frame information into the DeepSort tracker for trajectory tracking to obtain the vehicle target tracking trajectory includes the following steps:

Through the Kalman filter algorithm, the target position of the next frame is predicted based on the target motion trajectory extracted from the current input video;

Through the feature extraction network, the corresponding target appearance features are extracted according to the target position;

Determine a cost matrix for each detection frame feature and the target appearance feature according to the vehicle detection frame information, where the cost matrix represents the similarity between features;

Through the Hungarian algorithm, the target motion trajectory and the detection frame are associated according to the cost matrix to update the target motion trajectory.

According to some embodiments of the present invention, correcting the vehicle position coordinates in the vehicle target tracking trajectory according to camera parameters to obtain the corrected vehicle target tracking trajectory includes the following steps:

Using the distortion coefficient of the camera parameters, each image pixel of the surveillance video is subjected to distortion transformation to project to the camera normalized plane;

Project the image pixels after distortion transformation on the standardized plane to the pixel plane through the internal parameter matrix of the camera parameters, and obtain the pixel conversion relationship between the distorted pixels and the original image pixels on the pixel plane;

Each vehicle position coordinate in the vehicle target tracking trajectory is modified according to the pixel point conversion relationship to obtain a corrected vehicle target tracking trajectory.

On the other hand, embodiments of the present invention also provide a vehicle target tracking system, including:

The first module is used to obtain surveillance video;

The second module is used to input each frame of the surveillance video into the YOLOv7 detector in sequence for target detection to obtain vehicle detection frame information. The ELAN structure of the backbone network layer of the YOLOv7 detector is provided with a convolution block attention. Module, the convolution block attention module performs global maximum pooling and global average pooling operations on the input feature map to obtain an output feature map that fuses spatiotemporal attention features;

The third module is used to input each frame image and the vehicle detection frame information into the DeepSort tracker for trajectory tracking to obtain the vehicle target tracking trajectory;

The fourth module is used to correct the vehicle position coordinates in the vehicle target tracking trajectory according to the camera parameters to obtain the corrected vehicle target tracking trajectory.

On the other hand, embodiments of the present invention also provide a vehicle target tracking device, including:

at least one processor;

At least one memory for storing at least one program;

When the at least one program is executed by the at least one processor, at least one of the processors implements the vehicle target tracking method as described above.

On the other hand, embodiments of the present invention also provide a computer-readable storage medium that stores computer-executable instructions, and the computer-executable instructions are used to cause the computer to execute the vehicle target as described above. Tracking methods.

The above technical solution of the present invention has at least one of the following advantages or beneficial effects: Each frame of the surveillance video is input into the YOLOv7 detector in sequence for target detection to obtain vehicle detection frame information. The YOLOv7 detector is set in the ELAN structure of the backbone network layer. There is a convolution block attention module. The convolution block attention module performs global maximum pooling and global average pooling operations on the input feature map to obtain an output feature map that integrates spatiotemporal attention features. Through the convolution block in the ELAN structure The attention module is used to improve the representation ability of features, thereby improving the vehicle target detection accuracy of the YOLOv7 detector. Then input the vehicle detection frame information into the DeepSort tracker for trajectory tracking to obtain the vehicle target tracking trajectory, and then correct the vehicle position coordinates in the vehicle target tracking trajectory according to the camera parameters to obtain the corrected vehicle target tracking trajectory, which improves the accuracy of vehicle tracking. .

Description of drawings

Figure 1 is a flow chart of a vehicle target tracking method provided by an embodiment of the present invention;

Figure 2 is a schematic structural diagram of an ELAN provided by an embodiment of the present invention;

Figure 3 is a schematic diagram of the position of the convolution block attention module in the ELAN structure provided by the embodiment of the present invention;

Figure 4 is a schematic structural diagram of a vehicle target tracking device provided by an embodiment of the present invention.

Detailed ways

Embodiments of the present invention are described in detail below, examples of which are illustrated in the drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements with the same or similar functions. The embodiments described below with reference to the drawings are exemplary and are only used to explain the present invention and cannot be understood as limiting the present invention.

In the description of the present invention, it should be understood that the orientation descriptions involved, such as the orientations or positional relationships indicated by up, down, left, right, etc. are based on the orientations or positional relationships shown in the drawings, and are only for the convenience of describing the present invention. The invention and simplified description are not intended to indicate or imply that the devices or elements referred to must have a specific orientation, be constructed and operate in a specific orientation, and therefore are not to be construed as limitations of the present invention.

In the description of the present invention, if the first, second, etc. are described, they are only used for the purpose of distinguishing technical features, and cannot be understood as indicating or implying the relative importance or implicitly indicating the number or implicit indication of the indicated technical features. The sequence relationship of the indicated technical features.

Embodiments of the present invention provide a vehicle target tracking method, system, device and storage medium. The vehicle target tracking method in the embodiment of the present invention can be applied to terminals, can also be applied to servers, and can also be run on terminals. Or the software in the server, etc. The terminal can be a tablet computer, a laptop computer, a desktop computer, etc., but is not limited thereto. The server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers. It can also provide cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, and middleware. Cloud servers for basic cloud computing services such as software services, domain name services, security services, CDN, and big data and artificial intelligence platforms.

Referring to Figure 1, the vehicle target tracking method according to the embodiment of the present invention includes but is not limited to step S110, step S120, step S130 and step S140.

Step S110, obtain surveillance video;

Step S120, input each frame of the surveillance video into the YOLOv7 detector in turn for target detection to obtain vehicle detection frame information. Among them, the ELAN structure of the backbone network layer of the YOLOv7 detector is equipped with a convolution block attention module. The convolution block The attention module performs global maximum pooling and global average pooling operations on the input feature map to obtain an output feature map that fuses spatiotemporal attention features;

Step S130, input the vehicle detection frame information into the DeepSort tracker for trajectory tracking, and obtain the vehicle target tracking trajectory;

Step S140: Correct the vehicle position coordinates in the vehicle target tracking trajectory according to the camera parameters to obtain the corrected vehicle target tracking trajectory.

In some embodiments of step S120, YOLOv7 is a version of the YOLO model (You Only Look Once: Unified, Real-Time Object Detection, an object detection system based on a single neural network). The YOLO model is a deep learning algorithm that can be used for image recognition in computer vision technology. The YOLO model transforms the target detection problem into a Regression regression problem, that is, given an input image, it directly regresses the target's bounding box (bounding box) and its classification category at multiple locations in the image. YOLO models include but are not limited to Yolov3, Yolov4, Yolov5, Yolov7 (all different versions of YOLO), etc. Different models have different weights, network structure diagrams, and algorithms, and the regional sampling methods used are also different.

The overall network architecture of the yolov7 detector mainly consists of three parts: input (input layer), backbone (backbone network layer) and head (head layer). The backbone is used to extract features, and the head is used for prediction. The yolov7 detector first preprocesses the input image and then inputs it into the backbone network. According to the three-layer output in the backbone network, the head layer continues to output three layers of feature maps of different sizes through the backbone network. After RepVGG block and conv, the Predict the three types of image detection tasks (classification, front and rear background classification, and border) and output the final results. Among them, the backbone layer of yolov7 consists of several CBS modules, ELAN modules and MP modules. The CBS module is mainly used for convolution operations; the ELAN module is an efficient network structure that controls the shortest and longest gradient paths to make the network It can learn more features and has stronger robustness; the MP module is mainly used for downsampling operations.

In the embodiment of the present invention, referring to Figure 2, the improved ELAN structure includes a first branch, a second branch, a third branch and a fourth branch. The first branch and the second branch are both 1×1 convolution layers. The three branches are 1×1 convolution layer and 3×3 convolution layer, and the fourth branch is 1×1 convolution layer, 3×3 convolution layer and 3×3 convolution layer. . The first branch, the second branch, the third branch and the fourth branch are all connected to the splicing module. In the splicing module, the output of each branch is spliced using the Concat function. Specifically, modify the training configuration file yolov7.yaml of the YOLOv7 detector, modify the backbone part, delete a total of 8 convolutional layers, layers 7, 8, 20, 21, 33, 34, 46, and 47, and replace the 10th and 23rd convolutional layers. A total of four Concat layers, layers 36 and 49, are modified to [[-1,-2,-3,-4],1,Concat,[1]].

Further, referring to Figure 3, an improved convolution block attention module (CBAM) is added to the ELAN structure. The convolution block attention module is connected in series after the first convolution module (i.e., the first branch) in the optimized ELAN structure. The convolution block attention module performs a pooling operation on the features output by the convolution module based on the attention mechanism, so that the model can strengthen the representation of important features and improve the target detection accuracy of the model. The convolutional block attention module includes a channel attention mechanism unit and a spatial attention mechanism unit that are connected in sequence. The working processes of the channel attention mechanism unit and the spatial attention mechanism unit are as follows:

Channel attention mechanism unit: The input feature F (that is, the output feature of the first branch) is pooled in the channel dimension through the global maximum pooling layer and the global average pooling layer to obtain the channel global maximum pooling value and the channel global Average pooling value; the channel global maximum pooling value and the channel global average pooling value are respectively input into a layer of fully connected neural network and activated through the Relu activation function to obtain the first channel feature and the second channel feature; the first channel The feature and the second channel feature are element-wise based on the addition operation, and then through the sigmoid activation operation, the channel attention feature M _c is obtained; the channel attention feature M _c and the input feature F of the input channel attention mechanism unit are element-wise -wise multiplication operation to obtain the output feature F′ of the channel attention mechanism unit.

Spatial attention mechanism unit: perform global maximum pooling and global average pooling on the spatial dimension of the input features of the spatial attention mechanism unit (i.e., the output features F′ of the attention mechanism unit) to obtain the spatial global maximum pooling features. and spatial global average pooling features; after splicing the spatial global maximum pooling features and the spatial global average pooling features, the dimensionality is reduced to 1 channel through the convolution layer, and then through sigmoid activation, the spatial attention features are obtained; the spatial attention features are M _s and the input feature F′ of the input spatial attention mechanism unit perform element-wise multiplication operations to obtain the output features of the spatial attention mechanism unit, that is, the output feature map that fuses spatiotemporal attention features.

In some embodiments, the YOLOv7 detector is trained by a deep learning algorithm. During the training process of the YOLOv7 detector, it is necessary to calculate the model loss based on the predicted box and the real box output by the YOLOv7 detector, and continuously update the YOLOv7 detector based on the model loss. Until the output of the YOLOv7 detector reaches the preset accuracy. Specifically, the Wise-IoU loss function is used for calculation during the YOLOv7 detector training process, as follows:

According to the degree of overlap between the predicted box output by the YOLOv7 detector and the real box, the intersection and union ratio loss is determined. The intersection and union ratio loss is defined as:

L _IoU = 1-IOU;

Among them, IoU represents the intersection and union ratio of the predicted box and the real box.

The loss weight is determined based on the distance metric between the predicted box and the real box output by the YOLOv7 detector. The loss weight is defined as:

Among them, x and y are the coordinates of the center point of the prediction box, x _gt and y _gt are the coordinates of the center point of the real box, W _g and H _g represent the maximum width and height of the sum of the prediction box and the real box, and exp(·) represents the exponential operation.

Based on the attention mechanism, the model loss is determined based on the loss weight and the intersection-union ratio loss. The model loss is expressed as follows:

_LWIoU = _{RWIoULIoU ;}

Furthermore, after training, adjust the hyperparameters according to the target data set. Specifically, you can select the evolve option to train the data set again, train for 300 rounds, 10 generations per round, and find the optimal hyperparameter configuration.

In some embodiments of step S130, the DeepSort tracker is an algorithm that combines the two tasks of target detection and target tracking. After detecting the position and bounding box of the target object in each frame, the feature representation of the target is extracted through a deep learning model (such as CNN), and each target is matched with the target that has been tracked in the previous frame. During the matching process, factors such as target feature similarity and motion consistency will be considered to determine the target's identity and trajectory.

Specifically, in step S130, the step of inputting the vehicle detection frame information into the DeepSort tracker for trajectory tracking to obtain the vehicle target tracking trajectory includes but is not limited to the following steps:

Step S210, use the Kalman filter algorithm to predict the target position of the next frame based on the target motion trajectory extracted from the currently input video, where the target motion trajectory includes an uncertain state trajectory and a definite state trajectory;

Step S220: Extract corresponding target appearance features according to the target location through the feature extraction network;

Step S230, determine the cost matrix of each detection frame feature and the target appearance feature based on the vehicle detection frame information, where the cost matrix represents the similarity between the features;

Step S240, use the Hungarian algorithm to associate the target motion trajectory and the detection frame according to the cost matrix to update the target motion trajectory.

In this embodiment, the YOLOv7 detector of the embodiment of the present invention is used to identify vehicle targets in the surveillance video frame by frame to obtain the vehicle detection frame information of the vehicle target. The vehicle detection frame information includes but is not limited to coordinate information, category, and confidence. degree and image features, etc., and input the vehicle detection frame information of the vehicle target into the DeepSort tracker. In the DeepSort tracker, it is necessary to create the first batch of uncertain state trajectories based on the results of the first frame detection, and then through the correlation matching results between the detection frame and the uncertain state trajectory, and subsequent updates to obtain a batch of definite state trajectories. The DeepSort tracker process as follows:

S1. Perform IOU matching between the detection frame of the current frame and the target trajectory frame predicted by Kalman filtering based on the target motion trajectory of the previous frame and calculate the cost matrix. Enter the cost matrix into the Hungarian algorithm to obtain three matching results. First If the first type is trajectory mismatch, then delete the determined and uncertain trajectories in which the mismatch reaches the predicted number of times (such as 30 times); the second type is target mismatch, then create a new trajectory; the third type is trajectory matching, then Indicates that the tracking is successful, update the trajectory variables corresponding to the target through Kalman filtering, and repeat the current step until a definite trajectory frame appears or the video frame ends.

S2. Predict the definite state trajectory and the uncertain state trajectory through Kalman filtering. Perform cascade matching between the definite state trajectory and the target detection frame to obtain three matching results. The first one is trajectory matching, which is updated through Kalman filtering. ; The second result is trajectory mismatch; the third result is target mismatch. At this time, the previous uncertain trajectory, mismatched trajectory and mismatched target are IOU matched, and then the cost matrix is calculated through the matching results. .

S3. Input the cost matrix into the Hungarian algorithm and obtain three matching results. The first one is trajectory mismatch, then delete the definite state trajectory and uncertain state trajectory where the mismatch reaches 30 times. The second one is the target mismatch, then Create a new trajectory. The third type is trajectory matching, which means the tracking is successful. The trajectory variables corresponding to the target are updated through Kalman filtering.

S4. Repeat step S3 until the end of the video frame to obtain the vehicle target position and vehicle target trajectory.

According to some embodiments of the present invention, in step S140, the step of correcting the vehicle position coordinates in the vehicle target tracking trajectory according to the camera parameters to obtain the corrected vehicle target tracking trajectory includes but is not limited to the following steps:

Step S310, use the distortion coefficient of the camera parameters to perform distortion transformation on each image pixel of the surveillance video to project it to the camera standardized plane;

Step S320: Project the distorted image pixels on the standardized plane to the pixel plane through the internal parameter matrix of the camera parameters to obtain the pixel conversion relationship between the distorted pixels and the original image pixels on the pixel plane;

Step S330: Modify each vehicle position coordinate in the vehicle target tracking trajectory according to the pixel point conversion relationship to obtain a corrected vehicle target tracking trajectory.

In this embodiment, the internal parameter matrix and distortion parameters of the camera and the pixel size of the surveillance video are obtained, and then the pixel points on the video image are projected to the normalize plane of the camera, distortion transformation is performed on the normalize plane, and then the distorted points are projected Return to the pixel plane, obtain the corresponding conversion relationship between the distorted pixels and the original image pixels, repeat the above process until all pixel conversion relationships are obtained, and change the vehicle position coordinates in the trajectory output by the DeepSort tracker according to the pixel conversion relationship, and obtain the corrected vehicle target tracking trajectory.

According to some embodiments of the present invention, the improved YOLOv7 detector uses an optimized ELAN structure to replace the ELAN structure in the original YOLOv7 network architecture, adds an improved convolution attention module to the optimized ELAN structure, and uses the Wise-IoU loss function. Model training, and evaluating and tuning the hyperparameters of the training model based on the data set to improve the accuracy of vehicle target tracking from the surveillance perspective. By de-distorting the tracking trajectory coordinates, the accuracy of vehicle target motion trajectory extraction is improved.

An embodiment of the present invention also provides a vehicle target tracking system, including:

The first module is used to obtain surveillance video;

The second module is used to input each frame of the surveillance video into the YOLOv7 detector in sequence for target detection to obtain vehicle detection frame information. Among them, the ELAN structure of the backbone network layer of the YOLOv7 detector is equipped with a convolution block attention module. The convolution block attention module performs global maximum pooling and global average pooling operations on the input feature map to obtain an output feature map that fuses spatiotemporal attention features;

The third module is used to input each frame image and vehicle detection frame information into the DeepSort tracker for trajectory tracking to obtain the vehicle target tracking trajectory;

The fourth module is used to correct the vehicle position coordinates in the vehicle target tracking trajectory according to the camera parameters to obtain the corrected vehicle target tracking trajectory.

It can be understood that the contents in the above vehicle target tracking method embodiment are applicable to this system embodiment. The functions implemented by this system embodiment are the same as those in the above vehicle target tracking method embodiment, and the beneficial effects achieved are the same as those mentioned above. The beneficial effects achieved by the embodiments of the vehicle target tracking method are also the same.

Referring to Figure 4, Figure 4 is a schematic diagram of a vehicle target tracking device provided by an embodiment of the present invention. The vehicle target tracking device according to the embodiment of the present invention includes one or more control processors and memories. In FIG. 4 , one control processor and one memory are taken as an example.

The control processor and the memory can be connected through a bus or other means. Figure 4 takes the connection through a bus as an example.

As a non-transitory computer-readable storage medium, memory can be used to store non-transitory software programs and non-transitory computer executable programs. In addition, the memory may include high-speed random access memory and may also include non-transitory memory, such as at least one magnetic disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory may optionally include memory located remotely relative to the control processor, and these remote memories may be connected to the vehicle target tracking device via a network. Examples of the above-mentioned networks include but are not limited to the Internet, intranets, local area networks, mobile communication networks and combinations thereof.

Those skilled in the art can understand that the device structure shown in Figure 4 does not constitute a limitation on the vehicle target tracking device, and may include more or fewer components than shown, or combine certain components, or arrange different components. .

The non-transient software programs and instructions required to implement the vehicle target tracking method applied to the vehicle target tracking device in the above embodiment are stored in the memory, and when executed by the control processor, execute the vehicle target tracking method applied to the vehicle target tracking device in the above embodiment. Vehicle target tracking method.

In addition, one embodiment of the present invention also provides a computer-readable storage medium, which stores computer-executable instructions. The computer-executable instructions are executed by one or more control processors, which can cause the above-mentioned One or more control processors execute the vehicle target tracking method in the above method embodiment.

Those of ordinary skill in the art can understand that all or some steps and systems in the methods disclosed above can be implemented as software, firmware, hardware, and appropriate combinations thereof. Some or all of the physical components may be implemented as software executed by a processor, such as a central processing unit, a digital signal processor, or a microprocessor, or as hardware, or as an integrated circuit, such as an application specific integrated circuit . Such software may be distributed on computer-readable media, which may include computer storage media (or non-transitory media) and communication media (or transitory media). As is known to those of ordinary skill in the art, the term computer storage media includes volatile and nonvolatile media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. removable, removable and non-removable media. Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, Digital Versatile Disk (DVD) or other optical disk storage, magnetic cassettes, tapes, disk storage or other magnetic storage devices, or may Any other medium used to store the desired information and that can be accessed by a computer. Additionally, it is known to those of ordinary skill in the art that communication media typically embodies computer readable instructions, data structures, program modules or other data in a modulated data signal such as a carrier wave or other transport mechanism, and may include any information delivery media .

The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those of ordinary skill in the art, various modifications can be made without departing from the purpose of the present invention. Variety.

Claims (10)

1. A vehicle target tracking method, characterized in that it includes the following steps: Obtain surveillance video; Each frame of the surveillance video is input into the YOLOv7 detector in turn for target detection to obtain vehicle detection frame information. The ELAN structure of the backbone network layer of the YOLOv7 detector is provided with a convolution block attention module. The block attention module performs global maximum pooling and global average pooling operations on the input feature map to obtain an output feature map that fuses spatiotemporal attention features; Input the vehicle detection frame information into the DeepSort tracker for trajectory tracking to obtain the vehicle target tracking trajectory; The vehicle position coordinates in the vehicle target tracking trajectory are corrected according to the camera parameters to obtain a corrected vehicle target tracking trajectory. 2. The vehicle target tracking method according to claim 1, wherein the ELAN structure includes a first branch, a second branch, a third branch and a fourth branch, the first branch and the second branch Both are 1×1 convolutional layers. The third branch is successively a 1×1 convolutional layer and a 3×3 convolutional layer. The fourth branch is successively a 1×1 convolutional layer and a 3×3 convolutional layer. 3×3 convolution layer and 3×3 convolution layer; the first branch, the second branch, the third branch and the fourth branch are all connected to the splicing module. 3. The vehicle target tracking method according to claim 2, wherein the convolution block attention module is provided between the first branch and the splicing module, and the convolution block attention module includes The channel attention mechanism unit and the spatial attention mechanism unit are connected in sequence. 4. The vehicle target tracking method according to claim 3, characterized in that the channel attention mechanism unit is used for: The input features of the channel attention mechanism unit are pooled in the channel dimension respectively, and the channel global maximum pooling value and the channel global average pooling value are obtained; The channel global maximum pooling value and the channel global average pooling value are respectively input into the fully connected neural network and then activated to obtain the first channel feature and the second channel feature; The first channel feature and the second channel feature are added and activated to obtain the channel attention feature; Perform a multiplication operation on the channel attention feature and the input feature of the input channel attention mechanism unit to obtain the output feature of the channel attention mechanism unit; The spatial attention mechanism unit is used for: The input features of the spatial attention mechanism unit are pooled in the spatial dimension respectively to obtain the spatial global maximum pooling features and the spatial global average pooling features; After splicing the spatial global maximum pooling features and the spatial global average pooling features, perform dimensionality reduction and activation operations to obtain spatial attention features; The spatial attention feature and the input feature of the spatial attention mechanism unit are multiplied to obtain the output feature of the spatial attention mechanism unit. 5. The vehicle target tracking method according to claim 1, characterized in that the YOLOv7 detector is obtained through deep learning algorithm training, and the model loss during the training process of the YOLOv7 detector is obtained through the following steps: According to the degree of overlap between the predicted box output by the YOLOv7 detector and the real box, the intersection ratio loss is determined; Determine the loss weight based on the distance measure between the predicted box and the real box output by the YOLOv7 detector; Based on the attention mechanism, the model loss is determined based on the loss weight and the intersection-union ratio loss. 6. The vehicle target tracking method according to claim 1, characterized in that said inputting the vehicle detection frame information into the DeepSort tracker for trajectory tracking to obtain the vehicle target tracking trajectory includes the following steps: Through the Kalman filter algorithm, the target position of the next frame is predicted based on the target motion trajectory extracted from the current input video; Through the feature extraction network, the corresponding target appearance features are extracted according to the target position; Determine a cost matrix for each detection frame feature and the target appearance feature according to the vehicle detection frame information, where the cost matrix represents the similarity between features; Through the Hungarian algorithm, the target motion trajectory and the detection frame are associated according to the cost matrix to update the target motion trajectory. 7. The vehicle target tracking method according to claim 1, wherein correcting the vehicle position coordinates in the vehicle target tracking trajectory according to camera parameters to obtain the corrected vehicle target tracking trajectory includes the following steps: Using the distortion coefficient of the camera parameters, each image pixel of the surveillance video is subjected to distortion transformation to project to the camera normalized plane; Project the image pixels after distortion transformation on the standardized plane to the pixel plane through the internal parameter matrix of the camera parameters, and obtain the pixel conversion relationship between the distorted pixels and the original image pixels on the pixel plane; Each vehicle position coordinate in the vehicle target tracking trajectory is modified according to the pixel point conversion relationship to obtain a corrected vehicle target tracking trajectory. 8. A vehicle target tracking system, characterized by including: The first module is used to obtain surveillance video; The second module is used to input each frame of the surveillance video into the YOLOv7 detector in sequence for target detection to obtain vehicle detection frame information. The ELAN structure of the backbone network layer of the YOLOv7 detector is provided with a convolution block attention. Module, the convolution block attention module performs global maximum pooling and global average pooling operations on the input feature map to obtain an output feature map that fuses spatiotemporal attention features; The third module is used to input each frame of image and the vehicle detection frame information into the DeepSort tracker for trajectory tracking to obtain the vehicle target tracking trajectory; The fourth module is used to correct the vehicle position coordinates in the vehicle target tracking trajectory according to the camera parameters to obtain the corrected vehicle target tracking trajectory. 9. A vehicle target tracking device, characterized in that it includes: at least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, at least one of the processors implements the vehicle target tracking method according to any one of claims 1 to 7. 10. A computer-readable storage medium in which a processor-executable program is stored, wherein the processor-executable program, when executed by the processor, is used to implement any of claims 1 to 7. The vehicle target tracking method described in one item.