Real-time vehicle tracking method and device for road multi-camera view area limitation

Abstract

The invention provides a real-time vehicle tracking method and device for road multi-camera view area limitation, and the method comprises the steps: carrying out the multi-target tracking of a road vehicle under the view of each single camera, and obtaining the image coordinates of a vehicle center point in each frame; respectively defining a mask of a vehicle driving area for each view angle, and selecting a plurality of pairs of homonymy calibration points of the two view angles to solve a homography matrix; mapping a drivable area vertex of the first view angle to a second view angle coordinate system through the matrix to obtain a public drivable area; in the second view angle, only the vehicle center point falling into the area is reserved as a matching candidate, and reverse mapping is symmetrically executed to realize bidirectional area limitation; placing the center points of the limited two-view-angle vehicle in the same coordinate system, calculating the distance between the center points of the vehicle, and matching the same vehicle; and the tracking identifier of the same vehicle under the two view angles is updated according to the matching result, and real-time, efficient and accurate vehicle continuous tracking under multiple cameras is supported. The method is suitable for multi-camera real-time scene monitoring in an intelligent traffic system.

Description

Technical Field

[0001] The invention belongs to the field of intelligent traffic vehicle tracking and relates to a real-time vehicle tracking technical solution with limited field of view areas of multiple road cameras. Background Art

[0002] With rising living standards, cars have become a more and more common means of transportation. This increasing number of vehicles has also placed enormous pressure on urban management. With the rapid development of computer vision in recent years, it has become possible to detect, identify, and track vehicles using surveillance video. This allows for constant monitoring of vehicle information on roads, significantly reducing the labor costs of urban traffic management. Vehicle tracking is a crucial component of intelligent transportation systems. Based on captured video images, these systems can perform tasks such as traffic flow control, abnormal vehicle behavior detection, and pedestrian identification. Vehicle tracking is crucial for preventing and handling traffic accidents. Vehicle tracking ensures that a specific vehicle within its field of view can be automatically located, enabling it to be found in the event of an accident. For some traffic incidents, tracking data from a single camera is often insufficient. When tracking data covering a wider field of view is required, tracking results from multiple cameras must be matched to obtain a larger area of ​​vehicle activity. Currently, multi-camera vehicle tracking involves matching the paths tracked by a single camera. This matching process requires multiple frames of information from the same vehicle for feature extraction and network model training, which increases the complexity and computational effort of the matching process. Furthermore, the vehicle must complete the entire road section, making real-time tracking by multiple cameras impossible.

[0003] Therefore, this field urgently needs to propose new technical solutions to overcome the shortcomings of existing technologies, so as to break through the technical bottleneck of increased mismatching rate caused by the lack of bidirectional consistency verification of candidate regions for cross-view matching. Summary of the Invention

[0004] In response to the technical problems of low cross-view matching efficiency, high mismatching rate and difficulty in real-time operation in existing road multi-camera vehicle tracking algorithms, the purpose of the present invention is to provide a new real-time vehicle tracking technical solution with limited field of view of multiple road cameras. By utilizing the homography mapping of the drivable area of ​​each camera and the extraction of overlapping areas, efficient, accurate and real-time tracking of vehicles across viewpoints can be achieved.

[0005] The present invention provides a real-time vehicle tracking method with limited field of view of multiple cameras on a road, comprising the following steps: Under the field of view of each single camera, multiple vehicles on the road are tracked to obtain the image coordinates of the vehicle center point in each frame; A mask for the vehicle's drivable area is defined for each viewpoint. Multiple pairs of identically named calibration points from both views are selected to solve the homography matrix. This matrix is ​​used to map the vertices of the drivable area from the first viewpoint to the coordinate system of the second viewpoint, yielding a common drivable area. In the second viewpoint, only the center points of vehicles falling within this area are retained as matching candidates, and a symmetrical reverse mapping is performed to achieve bidirectional area restriction. Place the center points of the two restricted viewpoints in the same coordinate system, calculate the distance between the center points of the vehicles, and match the same vehicle based on the minimum distance principle; The tracking identification of the same vehicle in the two perspectives is updated according to the matching results, and the above steps are repeated for the new vehicle entering the field of view to achieve real-time continuous tracking.

[0006] Moreover, single-camera tracking is implemented using the DeepSort algorithm, and target detection during tracking uses the RetinaNet model independently trained for different perspective distortions.

[0007] Moreover, the homography matrix H in step b is a 3×3 homogeneous matrix, which is obtained by solving no less than four pairs of same-name points in the two-view images.

[0008] Moreover, when calculating the distance, all vehicles in the same frame are first paired, and candidate pairs that fall in the non-restricted area after mapping are eliminated.

[0009] Moreover, Euclidean distance is used to calculate the distance between vehicle center points.

[0010] Alternatively, when calculating the distance between vehicle center points, distance combined with weighted similarity is used.

[0011] Moreover, when updating the tracking identifier, the matching ID is merged, the new track is initialized, and the invalid track is destroyed.

[0012] On the other hand, the present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the real-time vehicle tracking method with the field of view area limitation of multiple cameras on the road as described above is implemented.

[0013] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the real-time vehicle tracking method with limited field of view of multiple cameras on the road as described above.

[0014] On the other hand, the present invention also provides a computer program product, including a computer program, which, when executed by a processor, implements the real-time vehicle tracking method with limited field of view areas of multiple road cameras as described above.

[0015] The present invention provides a real-time vehicle tracking method based on the limitation of the field of view areas of multiple road cameras. According to the characteristic that the field of view areas of road cameras overlap, the homography relationship between the fields of view of two cameras is calculated by using the same-name points in the overlapping area. The corresponding relationship between the field of view areas is obtained through the homography relationship between the two fields of view. On the basis of single-camera vehicle tracking, the field of view area limitation is utilized and the homography relationship between the fields of view is used to match vehicles traveling in the overlapping area of ​​the two fields of view, ultimately realizing a real-time, efficient and accurate multi-camera vehicle tracking solution.

[0016] Compared with the existing technology, the present invention significantly reduces the number of candidates for cross-view matching by mapping the homography of the drivable area of ​​each camera and extracting the common area. It also achieves high efficiency and accuracy in multi-camera vehicle tracking by combining minimum distance matching with trajectory management strategies. The overall system architecture optimization ensures real-time processing capabilities and is suitable for online continuous vehicle tracking in large-scale road monitoring networks. Verified by actual road scenarios, the present invention maintains high matching accuracy and real-time performance in complex environments such as dense vehicles and interference from large vehicles, significantly outperforming traditional cross-camera tracking solutions. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 1 is a flow chart of real-time vehicle tracking under the limitation of the field of view of multiple cameras on a road according to an embodiment of the present invention.

[0018] Figure 2 Schematic diagram of camera field of view limitation according to an embodiment of the present invention.

[0019] Figure 3 Schematic diagram of a homography matrix according to an embodiment of the present invention.

[0020] Figure 4 2 is a schematic diagram of multi-camera real-time vehicle tracking according to an embodiment of the present invention. DETAILED DESCRIPTION

[0021] In order to facilitate ordinary technicians in this field to understand and implement the present invention, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the implementation examples described herein are only used to illustrate and explain the present invention and are not used to limit the present invention.

[0022] This method applies bidirectional regional restrictions to the mapped vehicle center points. This bidirectional mapping constraint ensures that vehicles are matched only within the overlapping region, improving accuracy. This method primarily uses manual or predefined masks and homography mapping of each camera's drivable area to achieve strict screening within the overlapping region, significantly improving the efficiency and accuracy of cross-viewpoint vehicle matching.

[0023] The method includes the following steps: within the field of view of each single camera, using a combination of RetinaNet and DeepSort to detect and track road vehicles in real time and obtain the coordinates of the vehicle center point in each frame; using a manually outlined or pre-defined polygonal mask of the drivable area of ​​each viewpoint, and selecting ≥4 pairs of calibration points with the same name from two viewpoints to solve the homography matrix H, projecting the i-th viewpoint mask to the j-th viewpoint through H, and reversely mapping the j-th viewpoint mask to the i-th viewpoint, thereby extracting the common drivable area of ​​the two viewpoints; performing bidirectional regional restrictions on the mapped vehicle center points within each viewpoint, and retaining only points falling within the common area as cross-viewpoint matching candidates; mapping the filtered center points to the same coordinate system, calculating the Euclidean distance between the two viewpoint pairs, and completing vehicle matching based on the minimum distance principle; updating the matched vehicle tracking ID, and initializing the ID of the vehicle newly entering the field of view, and repeating the above process to achieve real-time, efficient, and accurate vehicle continuous tracking under multiple cameras. The present invention makes full use of the information of the overlapping field of view area, and through precise area restriction and mapping screening, greatly reduces the number of cross-camera matching candidates, reduces the computational complexity, and improves the matching robustness. It is suitable for multi-camera real-time monitoring scenarios in intelligent transportation systems.

[0024] The environment used in the present invention can be set according to the specific implementation scenario. For the convenience of implementation reference, the following embodiment environment description is provided: the operating system used is Ubuntu 16.04, the GPU model is NVIDIA Tesla M60 with 8 GB of video memory, and the CPU model is Intel (R) Xeon (R) CPU E5-2687W v4 @ 3.00GHz.

[0025] See also Figure 1 The embodiment of the present invention provides a real-time vehicle tracking method based on the field of view of multiple cameras on the road, including the following process: Step a: Track each vehicle on the road using a single camera In the field of view of each single camera, a detection-based multi-target tracking algorithm is used to track road vehicles and obtain the image coordinates of the center point of each vehicle in each frame.

[0026] This method is based on single-camera vehicle tracking. In the field of view of each camera, it is preferably recommended to use the single-camera target tracking algorithm DeepSort to track the vehicle separately to obtain the tracking results of each frame. And the tracked target is positioned, that is, the coordinates of the center point of each tracked vehicle relative to the image are obtained. The present invention further proposes to use a target detection model that is independently trained for different perspective distortions. The DeepSort algorithm is a typical detection-based multi-target tracking algorithm. The target detection network used in this method is RetinaNet. Due to the different degrees of distortion caused by imaging from different camera perspectives, in order to achieve better tracking effects, it is necessary to use different data sets for training camera perspectives with different distortions to obtain a deep learning network model suitable for target detection under different perspectives. In practice, the RetinaNet model can locate the positions of all vehicles within a video frame and output the confidence level of the vehicle detection boxes. Based on RetinaNet's detection results, the DeepSort algorithm correlates vehicles across frames, generating stable tracking trajectory center point coordinates. RetinaNet is the perception layer foundation of DeepSort, and DeepSort is RetinaNet's timing extender. These two are serially coupled to form a complete single-view tracking solution, supporting the multi-camera cross-view matching architecture proposed in this paper.

[0027] Step b: limit the location of the vehicle by area A polygonal mask of the vehicle's drivable area is defined for each perspective; multiple pairs of calibration points with the same name from the two perspectives are selected to solve the homography matrix; the vertices of the drivable area of ​​the first perspective are mapped to the coordinate system of the second perspective through the matrix to obtain a common drivable area; in the second perspective, only the center points of the vehicles falling into the area are retained as matching candidates, and the reverse mapping is performed symmetrically to achieve two-way area restriction, that is, the drivable area mask of the second perspective is reversely mapped to the first perspective to achieve two-way consistency restriction.

[0028] Furthermore, the homography matrix is ​​a 3×3 homogeneous matrix, which is obtained by solving no less than four pairs of same-name points in the two-view images.

[0029] Specifically, this can be achieved using the following process: 1) For each camera view, obtain a polygonal mask of the vehicle’s drivable area by manually outlining or pre-defining; 2) Select at least 4 pairs of identically named calibration points from the two camera perspectives and use the least squares method (which can be combined with RANSAC verification) to solve the 3×3 homography matrix H; 3) Map the polygonal vertices of the drivable area of ​​the i-th perspective to the j-th perspective coordinate system via H to obtain the common drivable area of ​​the two perspectives; 4) In the j-th view, only the vehicle center points that fall within the common area are retained as cross-view matching candidates; similarly, the j-th view mask is mapped back to the i-th view to achieve bidirectional consistency constraints.

[0030] In the embodiment, the position of the tracking vehicle obtained in step a is obtained by using the field of view area restriction map ( Figure 2 ) Restrictions are imposed on tracked vehicles. Figure 2 The vehicle travel areas in View 1 and View 2 can be initially outlined using software tools or manually. By observing the location of the tracked vehicle's center point, the restricted area is drawn. The area corresponding to the vehicle's travel location in View 2 in View 1 is mapped from View 2 to View 1 using the homography matrix H corresponding to the transformation relationship between the two camera perspectives. Ultimately, the tracked vehicle is matched within the constraints of the vehicle's travel area and the common view area.

[0031] The homography matrix represents the transformation relationship between the same-name points on the image obtained from two perspectives. On a two-dimensional image, the size of the homography is 3×3. Figure 3 shown.

[0032] For object A in the real three-dimensional world, the coordinates on the images obtained from perspective 1 and perspective 2 are , , then the homography matrix can be obtained ,in and Respectively , Homogeneous coordinates. For the solution of the homography matrix, a pair of points with the same name satisfy (1) in 、 Respectively represent the homogeneous coordinates of the same-name points under two perspectives, where 、 , it can be inferred from the above formula: (2) (3) Two equations can be obtained. The homography matrix contains 9 unknowns H1-H9. By finding more than four pairs of identical points in the two images, the homography matrix of the two image objects can be obtained.

[0033] By restricting the area of ​​tracked vehicles, it can be ensured that in the subsequent multi-camera vehicle matching process, the vehicle in perspective 1 can find its corresponding vehicle in perspective 2, which can reduce the computational complexity of the matching process and improve the matching accuracy.

[0034] Step c: Match the vehicles from two perspectives In this step, the center points of the two vehicle views restricted in step b are placed in the same coordinate system, and the Euclidean distance between the center points of each vehicle in the two views is calculated, and the same vehicle is matched based on the principle of minimum distance.

[0035] Furthermore, when calculating the Euclidean distance, all vehicles in the same frame are first paired, and candidate pairs that fall within the non-restricted area after mapping are eliminated to improve the matching accuracy.

[0036] In this embodiment, after obtaining the tracked vehicles from both perspectives in step b, the position of the vehicle in perspective 1 is transformed using the inter-view homography matrix. The center coordinates of the vehicles in both perspectives are now in the same coordinate system. The Euclidean distance between the tracked vehicle in perspective 1 and the vehicle in perspective 2 is then calculated. The vehicle with the shortest Euclidean distance between the two perspectives is considered the same vehicle. Due to the previously defined field of view restrictions, the corresponding vehicle in perspective 1 can always be found in perspective 2. This completes vehicle matching for multiple camera perspectives.

[0037] Step d: Update the ID of the vehicle matched under the two perspectives This step updates the tracking ID of the same vehicle in both viewpoints based on the matching results. Steps a–d are repeated continuously for each new vehicle that enters the field of view, achieving real-time vehicle tracking under multiple cameras.

[0038] Specifically, tracking ID updates and trajectory management are performed, including merging or updating the tracking IDs of the matching pairs under their respective perspectives based on the matching results; initializing new IDs for newly entered and unmatched vehicles; and automatically destroying trajectories that have not been matched for multiple frames and have left the public area. Steps a–d are then repeated to achieve real-time continuous vehicle tracking in a multi-camera network.

[0039] In this embodiment, the matched vehicle ID is modified based on the matching result obtained in step c. At this point, the matching of the current frame is completed, and the matching result for one frame is obtained. The relationship between the corresponding vehicles in the two perspectives can be linked, and then the matching of the newly entered vehicle is performed. In specific implementation, each camera captures road video in real time, and each frame is fed into a pre-trained multi-target detection and tracking network framework. Through detection and tracking, the detection frame, tracking ID, and center point coordinates of all vehicles in each frame are obtained in real time. Cross-perspective matching and ID update are completed according to the following process: Single-camera tracking: Target detection and DeepSort tracking are performed independently on each camera image, outputting the vehicle center coordinates and initial ID. Area mapping and screening: Use manually outlined or predefined drivable area polygon masks and map them with the homography matrix H to obtain the common drivable area from both perspectives, and impose bidirectional area restrictions on the vehicle center point. Cross-view matching: Map the filtered center points to the same coordinate system, calculate the Euclidean distance, and match the same vehicle based on the minimum distance principle (or weighted similarity); ID management: Merge the tracking IDs from each perspective in the matching pair, initialize the ID of the newly entered vehicle, and destroy the tracks that have not been matched in multiple frames and have left the public area. Repeat the above steps and output the continuous vehicle tracks under multiple cameras in real time.

[0040] For vehicle trajectories that have not been successfully matched for multiple consecutive frames and have moved out of the public drivable area, their tracking identification will be automatically terminated.

[0041] Furthermore, after each matching is completed, the matching results of adjacent frames may be smoothed based on the matching history to reduce jitter of the tracking ID. The present invention proposes a method for multi-camera vehicle tracking based on the overlapping area of ​​the camera field of view, which can realize multi-camera vehicle tracking efficiently, accurately and in real time. The method first uses the overlapping area to obtain the conversion relationship between the imaging under the perspectives of the two cameras, and then, on the basis of single-camera vehicle tracking, limits the field of view area to ensure matching accuracy and improve calculation efficiency, and uses the tracked target position information to match the vehicles under the two fields of view, and finally achieves the effect of multi-camera vehicle real-time tracking. In addition, in order to evaluate the experimental effect, the present invention proposes a successful matching rate RSM to evaluate the matching effect of the same vehicle between cameras. Experiments show that the method of the present invention has good matching effects in road scenes with a large number of vehicles, scenes with large vehicles such as buses, and road scenes with sudden appearance of vehicles, and the algorithm is efficient and meets the requirements of real-time tracking. The method proposed by the present invention will be applied to intelligent transportation systems to detect urban traffic. Reference Figure 4 , the application test experiment of the embodiment of the present invention is as follows: The present invention provides experimental data to demonstrate the effectiveness of the present invention. The experimental results are as follows: Figure 4 As shown, in order to further objectively evaluate the effect of multi-camera vehicle tracking, the present invention proposes a successful matching rate To quantitatively measure the accuracy of the algorithm in multi-camera vehicle tracking. Indicates the successful matching rate of the same target under two cameras based on single-camera multi-target tracking. It is calculated as follows:

[0042] in Indicates the number of vehicles that appear in both camera fields of view. Indicates the number of times the algorithm records the same vehicle ID under two cameras.

[0043] Evaluation of the experimental results revealed that a total of 73 vehicles passed through the fields of view of both cameras in the video data, and the single-camera tracking algorithm tracked all of them. Vehicles were matched in the overlapping area of ​​the two cameras, and a total of 66 vehicles were successfully matched, meaning the same vehicle had the same tracking ID in both fields of view. This resulted in an RSM success rate of 0.94 for the multi-camera vehicle tracking algorithm. Under the same configuration described above, the frame rate of the vehicle tracking algorithm for a single camera was 11.83 fps. Simultaneously tracking and matching the video data from both cameras yielded a final frame rate of 7.95 fps. Multi-camera vehicle tracking is a double-take of the single-camera vehicle tracking task, requiring twice the computational effort and memory as single-camera tracking. Furthermore, matching calculations for targets from both viewing angles are also required. However, the calculated frame rate is maintained, and the time required to perform the multi-camera vehicle tracking task is not twice that of the single-camera task, indicating that the algorithm has higher computational efficiency. The present invention matches vehicles in each frame of tracking, requiring only the position information of the current frame. Therefore, real-time multi-camera vehicle tracking is possible.

[0044] In specific implementation, the method proposed in the technical solution of the present invention can be automatically run by those skilled in the art using computer software technology. System devices that implement the method, such as computer-readable storage media that store the corresponding computer program of the technical solution of the present invention and computer equipment that runs the corresponding computer program, should also be within the scope of protection of the present invention.

[0045] The technical solution of the present invention can be automatically operated by those skilled in the art using a combination of computer software and hardware. In some possible embodiments, the real-time vehicle tracking system with limited field of view of multiple cameras on a road provided by the present invention may include: Image acquisition module: connects to multiple surveillance cameras, collects video streams from each camera, and is responsible for video stream acquisition and timestamp synchronization; Single-camera tracking module: performs deep learning-based detection and tracking, generating the center point coordinates and initial tracking ID of the vehicle in each frame; Area mapping module: stores the drivable area mask of each camera, and realizes polygon mask mapping and overlapping area extraction based on the homography matrix, and screens matching candidates; Cross-view matching module: For the center points that fall into the common area after mapping, it calculates the Euclidean distance and matches the minimum distance, or uses weighted similarity to match the center points of the vehicles, and outputs the matching pairs; Track management module: completes the merging, initialization, and invalidation of tracking IDs, including merging and updating cross-view IDs, initializing new tracks, and destroying invalid tracks; In some possible embodiments, the real-time vehicle tracking system with limited road multi-camera field of view provided by the present invention can be further divided into the following functional modules: Single-camera multi-target tracking module: realizes vehicle detection and tracking in various video streams; Drivable area definition and homography calibration module: supports manual or semi-automatic calibration of homonymous points, calculates the homography matrix, and generates drivable area masks for each viewpoint; Common area extraction and center point screening module: maps polygon masks, extracts overlapping areas, and screens matching candidates; Cross-view matching and similarity calculation module: completes cross-camera center point matching based on Euclidean distance (or distance + appearance similarity weighting); Tracking ID management module: performs ID merging, track initialization and expiration destruction of matching results; System communication and scheduling module: responsible for data scheduling, network communication and timing management between modules.

[0046] The following embodiments describe the electronic device provided by the present invention. The electronic device described below and the above-described method for real-time vehicle tracking with limited field of view of multiple cameras on a road can be referred to in correspondence with each other.

[0047] The electronic device may include a processor, a communications interface, a memory, and a communications bus, wherein the processor, the communications interface, and the memory communicate with each other via the communications bus. The processor may invoke logic instructions in the memory to execute a real-time vehicle tracking method with restricted field of view of multiple road cameras, primarily including the software processing portion of the aforementioned steps.

[0048] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage media include various media capable of storing program code, such as USB flash drives, mobile hard drives, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks.

[0049] On the other hand, an embodiment of the present invention also provides a computer program product, which includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the software processing part of the real-time vehicle tracking method with road multi-camera field of view area limitation provided by the above methods.

[0050] On the other hand, an embodiment of the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the software processing portion of the real-time vehicle tracking method with road multi-camera field of view area limitation provided by the above-mentioned methods.

[0051] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0052] Through the description of the above embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, 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, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0053] 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 various embodiments of the present invention.

Claims

1. A real-time vehicle tracking method with limited field of view of multiple cameras on a road, characterized in that: Including the following process, Under the field of view of each single camera, multiple vehicles on the road are tracked to obtain the image coordinates of the vehicle center point in each frame; A mask for the vehicle's drivable area is defined for each viewpoint. Multiple pairs of identically named calibration points from both views are selected to solve the homography matrix. This matrix is ​​used to map the vertices of the drivable area from the first viewpoint to the coordinate system of the second viewpoint, yielding a common drivable area. In the second viewpoint, only the center points of vehicles falling within this area are retained as matching candidates, and a symmetrical reverse mapping is performed to achieve bidirectional area restriction. Place the center points of the two restricted viewpoints in the same coordinate system, calculate the distance between the center points of the vehicles, and match the same vehicle based on the minimum distance principle; The tracking identification of the same vehicle in the two perspectives is updated according to the matching results, and the above steps are repeated for the new vehicle entering the field of view to achieve real-time continuous tracking.

2. The real-time vehicle tracking method with field of view restriction of multiple cameras on a road according to claim 1, characterized in that: Single-camera tracking is implemented using the DeepSort algorithm, and target detection during tracking uses the RetinaNet model independently trained for different perspective distortions.

3. The real-time vehicle tracking method with field of view restriction of multiple cameras on a road according to claim 1, characterized in that: The homography matrix H in step b is a 3×3 homogeneous matrix, which is obtained by solving no less than four pairs of same-name points in the two-view images.

4. The method for real-time vehicle tracking with multi-camera field of view restriction on a road according to claim 1, characterized in that: When calculating the distance, all vehicles in the same frame are first paired, and candidate pairs that fall into the non-restricted area after mapping are eliminated.

5. The real-time vehicle tracking method with road multi-camera field of view limitation according to claim 1 is characterized in that: When calculating the distance between vehicle center points, Euclidean distance is used.

6. The real-time vehicle tracking method with road multi-camera field of view restriction according to claim 1, characterized in that: When calculating the distance between vehicle center points, distance combined with weighted similarity is used.

7. The real-time vehicle tracking method with road multi-camera field of view limitation according to claim 1, characterized in that: When updating the tracking ID, merge the matching ID, initialize the new track and destroy the invalid track.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the real-time vehicle tracking method with road multi-camera field of view area limitation as described in any one of claims 1 to 7 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the real-time vehicle tracking method with road multi-camera field of view area limitation as claimed in any one of claims 1 to 7 is implemented.

10. A computer program product comprising a computer program, characterized in that: When the computer program is executed by a processor, the real-time vehicle tracking method with road multi-camera field of view area limitation as claimed in any one of claims 1 to 7 is implemented.

Images