A multi-target tracking method based on a linear camera model

Abstract

The application claims a multi-target tracking method based on a linear camera model. The method first performs target detection on the input video sequence, and performs lens distortion compensation on the detection result; then based on the linear camera model, the detection result in the pixel coordinate system is inversely projected onto the ground plane in the three-dimensional world coordinate system to obtain the measurement parameters of the target in the real world coordinate system and the uncertainty description thereof; further, a camera motion compensation mechanism is introduced to correct the errors caused by camera jitter or displacement; on this basis, the Kalman filter is used to predict and update the target state; finally, a similarity calculation method combining the projection Mahalanobis distance and the intersection over union (IoU) is proposed to realize more robust and accurate target association. The application effectively reduces the tracking errors caused by occlusion, camera motion and nonlinear motion in a complex dynamic scene, and significantly improves the stability and accuracy of multi-target tracking.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and video intelligent analysis technology, and in particular relates to a multi-target tracking method suitable for complex dynamic scenes. Background Technology

[0002] Multi-object tracking is a key research area in computer vision, aiming to continuously and stably locate and maintain the identity of multiple moving targets in a video sequence. Existing multi-object tracking methods mostly adopt a "detect first, track later" paradigm, typically relying on motion models or appearance features in the target association stage. However, in complex scenes, target occlusion, nonlinear motion, and camera jitter or displacement can significantly degrade tracking performance.

[0003] Traditional pixel-based motion modeling methods struggle to accurately reflect the motion of targets in the real world, especially under dynamic camera conditions where there may be no overlap between detection and prediction boxes, causing cross-union (CUI) based association methods to fail. Therefore, a multi-target tracking method that can model target motion in a real-world coordinate system and effectively eliminate camera motion interference is urgently needed.

[0004] A search revealed application CN115035154A, a visual multi-object tracking method based on deep learning object detection, applicable to the field of computer vision tracking. This method combines Kalman filtering and deep learning. First, a deep learning network model is used to detect objects in the image. Then, another deep learning network model is used to extract the depth feature map of the target. By comparing the depth feature maps of two frames, the historical trajectory of the target is matched. For targets that fail to match, a Kalman filter is used again for matching. If the match is successful, the target trajectory is updated; otherwise, it is considered a new target. This invention combines Kalman filtering tracking with deep learning, enabling fast and accurate target tracking. It solves the problems of slow response and large tracking errors in traditional methods, improving the robustness and accuracy of multi-object tracking in complex scenes. This patent performs motion modeling and correlation measurement on the pixel plane, but it has the following shortcomings: When the target distance changes or perspective scaling occurs, the pixel coordinates and box scale change significantly, which may lead to incorrect correlation; it lacks explicit compensation for lens distortion and camera motion: In complex scenarios such as vehicle-mounted vehicles, camera shake, turning, and lens distortion can introduce systematic measurement biases, causing inconsistencies between prediction and observation, and further reducing correlation robustness; the correlation strategy relies on empirical thresholds: Under conditions of occlusion, dense targets, and similar appearance, relying solely on appearance thresholds or IoU is prone to ID switching, making it difficult to balance accuracy and stability.

[0005] To address the aforementioned issues, this invention proposes a multi-target tracking method based on a linear camera model. The core of this method lies in: projecting image measurements from the pixel plane to world coordinates via inverse projection of the linear camera model for unified modeling and association, reducing the impact of perspective and scale changes on data association; introducing lens distortion compensation and camera motion compensation to eliminate systematic imaging errors and pseudo-displacements caused by self-motion before target association, improving observation consistency and trajectory continuity; and constructing a fusion of projected Mahalanobis distance and IoU based on Kalman filter prediction for similarity calculation, thereby significantly reducing mismatches and ID switching in occluded, dense, and visually similar scenes. Summary of the Invention

[0006] The purpose of this invention is to overcome the problem of insufficient robustness of existing multi-target tracking methods in complex dynamic scenes. It proposes a multi-target tracking method based on a linear camera model, which achieves more accurate and stable target association by inversely projecting the target motion model from pixel space to a three-dimensional world coordinate system.

[0007] The technical solution of the present invention is as follows:

[0008] A multi-target tracking method based on a linear camera model includes the following steps:

[0009] Step 1: Obtain the input video sequence, perform target detection on each frame, and obtain the detection results of the target in the pixel coordinate system;

[0010] Step 2: Perform lens distortion compensation on the detection results obtained in Step 1 to correct the imaging error caused by nonlinear distortion of the camera lens;

[0011] Step 3: Based on the linear camera model, the detection results after lens distortion compensation are inversely projected from the pixel coordinate system to the ground plane in the three-dimensional world coordinate system to obtain the inverse projection related measurement parameters of the target;

[0012] Step 4: After adding the relevant measurement parameters to the camera motion compensation, update the Kalman filter;

[0013] Step 5: Based on the similarity calculation method of projection Mahalanobis distance and intersection-union ratio fusion, the similarity between the prediction result and the detection result is calculated to obtain the tracking trajectory.

[0014] Furthermore, step 1 involves using a YOLOX detector (a single-stage target detection neural network used to infer the target's category, confidence level, and bounding box coordinates from the input video frame) to detect the target in the input video and obtain the detection result;

[0015] Furthermore, step 2: lens distortion compensation includes radial distortion compensation and tangential distortion compensation; specifically:

[0016] The mathematical model for radial distortion is:

[0017]

[0018] In the formula, The coordinates of the radial distortion point in the two-dimensional image plane coordinate system. The coordinates of the projection point on the two-dimensional image plane coordinate system. , , ... are the radial distortion coefficients corresponding to each order, respectively. This represents the distance between the ideal, distortion-free point and the principal point.

[0019] Tangential distortion is caused by asymmetry in lens assembly. The mathematical model for tangential distortion is as follows:

[0020]

[0021] In the formula, These are the coordinates of the tangential distortion point in the two-dimensional image plane coordinate system. , ... are the tangential distortion coefficients corresponding to each order;

[0022] The distortion model of a camera lens is represented as follows:

[0023]

[0024] In the formula, These are the coordinates in the two-dimensional image plane coordinate system under the distortion model.

[0025] Furthermore, step 3 involves inversely projecting the pixel coordinate system onto the three-dimensional world coordinate system to obtain the distribution of relevant measurement parameters in the three-dimensional world coordinate system; specifically:

[0026] According to the linear camera model, the transformation relationship between the world coordinate system and the pixel coordinate system is as follows:

[0027]

[0028] In the formula, As a scaling factor, In 3D camera coordinates A point on the axis; These are the coordinates in the pixel coordinate system. , ,in, Focal length and They represent direction and Focal length in direction, and They represent direction and Pixel density in the direction; Principal point coordinates (coordinates in the pixel coordinate system corresponding to the origin of the two-dimensional image plane coordinate system); For rotation matrix, It is a translation vector; A point in a three-dimensional world coordinate system; It is an intrinsic matrix that contains the camera's internal parameters. This is the external matrix, which contains the camera's external parameters; Let be the projection matrix from the 3D world coordinate system to the pixel coordinate system, where is used for ease of subsequent calculations. ~ express matrix.

[0029] make ,in, Let be a constant, and assume all target points lie in the same known depth plane. The above equation can be expressed as:

[0030]

[0031] make:

[0032]

[0033] In the formula, for ease of subsequent calculations, we use... ~ express matrix.

[0034] Therefore, the projection matrix of the pixel coordinate system inversely projected onto the three-dimensional world coordinate system is obtained. Combining the two equations above, we can obtain the linear relationship between the inverse projection of the pixel coordinate system to the three-dimensional world coordinate system:

[0035]

[0036] Through linear transformation, an approximate relationship for error propagation is obtained:

[0037]

[0038] In the formula, and These represent the errors in 3D world coordinates and pixel coordinates, respectively. For Jacobian matrices, Represented as:

[0039]

[0040] Assuming the detector's measurement error in the pixel coordinate system follows a Gaussian distribution, the measurement noise... Represented as:

[0041]

[0042] In the formula, Represents the noise factor. and These represent the width and height of the target, respectively.

[0043] Therefore, the covariance matrix of the measurement error in the three-dimensional world coordinate system is:

[0044]

[0045] Thus, the measurement noise is obtained by inversely projecting it from the pixel coordinate system to the three-dimensional world coordinate system.

[0046] Furthermore, step 4 specifically includes: extracting feature points in the image background, then applying the LK optical flow method to track the motion of the feature points between two frames; after obtaining the correspondence between the feature points in the two frames, calculating the affine transformation matrix of the image background motion using the RANSAC algorithm; finally, applying the affine transformation matrix to perform camera motion compensation on the Kalman filter prediction results; specifically:

[0047] RANSAC calculation of affine transformation matrix:

[0048]

[0049] In the formula, It refers to the scaling and rotation components of an affine transformation. This is the translation part; the camera motion compensation for the predicted state and prediction covariance of the Kalman filter is calculated as follows:

[0050]

[0051]

[0052]

[0053] in, and for and The extended matrix; and These are the predicted state vectors of the Kalman filter before and after camera motion compensation, respectively. and These are the prediction covariance matrices of the Kalman filter before and after correction, respectively; after this, the update steps of the Kalman filter become:

[0054]

[0055] In the formula, For Kalman gain, The observation matrix; For updated state estimation, These are measured values; For the updated covariance estimate, It is an identity matrix.

[0056] Furthermore, step 5 involves a similarity calculation method that integrates Mahalanobis distance and IoU. This method incorporates measurement noise projected onto the ground plane into the Mahalanobis distance calculation and adds IoU for auxiliary calculation; specifically:

[0057] Calculate the residuals:

[0058]

[0059] In the formula, For measured values, The observation matrix;

[0060] Calculate the residual covariance matrix:

[0061] ;

[0062] Therefore, the projected Mahalanobis distance can be calculated:

[0063] ;

[0064] use Impossible associations are excluded by calculating the threshold of the projected Mahalanobis distance for the 95% confidence interval; for a two-dimensional measurement space, the corresponding projected Mahalanobis distance threshold is... ;

[0065] By combining projected Mahalanobis distance with IoU, a new similarity calculation method is obtained:

[0066]

[0067] in, Indicates the first The predicted value, the first One measurement value; Here, the weighting coefficients are used to balance the contributions of the projected Mahalanobis distance and IoU to the final cost, let... .

[0068] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor, when executing the program, implements a multi-target tracking method based on a linear camera model as described in any one of the claims.

[0069] A non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the multi-target tracking method based on a linear camera model as described in any one of the claims.

[0070] The advantages and beneficial effects of this invention are as follows:

[0071] This invention has three main innovations: 1. Existing multi-target tracking algorithms primarily rely on pixel-plane correlation, failing to reflect target motion in the real-world plane. By using a linear camera model to inversely project pixel coordinates to 3D world coordinates, the target motion modeling becomes more consistent with real physical space (step 3); 2. Traditional tracking often treats detection boxes as "direct observations," rarely incorporating systemic biases caused by camera distortion into the source of correlation errors. Furthermore, in dynamic camera scenarios, camera shake causes significant discrepancies between the actual target motion and the predicted value, leading to mismatches between detection and prediction boxes, resulting in ID switching and tracking errors. By introducing lens distortion compensation and camera motion compensation, interference from imaging distortion and camera shake is effectively eliminated (steps 2 and 4); 3. Common similarity calculation methods rely on IoU matching, but IoU algorithms are based only on the overlapping area of ​​bounding boxes, without considering measurement noise and the uncertainty of measured and predicted values ​​in complex scenes. A similarity calculation method using projection Mahalanobis distance and intersection-union ratio (IoU) fusion improves the accuracy and robustness of target correlation (step 5).

[0072] This invention is applicable to multi-target tracking tasks in complex dynamic scenarios and has strong versatility and engineering practical value. Attached Figure Description

[0073] Figure 1 This is a flowchart of a preferred embodiment of the present invention for a multi-target tracking method based on a linear camera model. Detailed Implementation

[0074] The technical solutions of the embodiments of the present invention will be clearly and thoroughly described below with reference to the accompanying drawings. The described embodiments are merely some embodiments of the present invention.

[0075] The technical solution of the present invention to solve the above-mentioned technical problems is:

[0076] It should be noted that the user information (including but not limited to user device information, personal user information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the laws, regulations and standards of relevant countries and regions.

[0077] Please see Figure 1 This embodiment provides a multi-target tracking method based on a linear camera model, which includes the following steps:

[0078] Step 1: Apply the YOLOX detector (a single-stage object detection neural network used to infer the target's category, confidence score, and bounding box coordinates from the input video frame) to perform object detection on the input video and obtain the detection results;

[0079] Step 2: Based on the lens distortion model, lens distortion compensation includes radial distortion compensation and tangential distortion compensation; specifically:

[0080] Radial distortion is caused by changes in surface curvature due to optical defects during lens manufacturing. The mathematical model for radial distortion is:

[0081]

[0082] In the formula, The coordinates of the radial distortion point in the two-dimensional image plane coordinate system. The coordinates of the projection point on the two-dimensional image plane coordinate system. , , ... are the radial distortion coefficients corresponding to each order, respectively. This represents the distance between the ideal, distortion-free point and the principal point.

[0083] Tangential distortion is caused by asymmetry in lens assembly. The mathematical model for tangential distortion is as follows:

[0084]

[0085] In the formula, These are the coordinates of the tangential distortion point in the two-dimensional image plane coordinate system. , , ... are the tangential distortion coefficients corresponding to each order.

[0086] For most commonly used cameras, second-order distortion parameters are sufficient to meet accuracy requirements. Therefore, the distortion model of a camera lens can be expressed as:

[0087]

[0088] In the formula, These are the coordinates in the two-dimensional image plane coordinate system under the distortion model.

[0089] Step 3: Inversely project the pixel coordinate system onto the 3D world coordinate system to obtain the distribution of relevant measurement parameters in the 3D world coordinate system; specifically:

[0090] According to the linear camera model, the transformation relationship between the world coordinate system and the pixel coordinate system is as follows:

[0091]

[0092] In the formula, As a scaling factor, In 3D camera coordinates A point on the axis; These are the coordinates in the pixel coordinate system. , ,in, Focal length and They represent direction and Focal length in direction, and They represent direction and Pixel density in the direction; Principal point coordinates (coordinates in the pixel coordinate system corresponding to the origin of the two-dimensional image plane coordinate system); For rotation matrix, It is a translation vector; A point in a three-dimensional world coordinate system; It is an intrinsic matrix that contains the camera's internal parameters. This is the external matrix, which contains the camera's external parameters; Let be the projection matrix from the 3D world coordinate system to the pixel coordinate system, where is used for ease of subsequent calculations. ~ express matrix.

[0093] make ,in, Let be a constant, and assume all target points lie in the same known depth plane. The above equation can be expressed as:

[0094]

[0095] make:

[0096]

[0097] In the formula, for ease of subsequent calculations, we use... ~ express matrix.

[0098] Therefore, the projection matrix of the pixel coordinate system inversely projected onto the three-dimensional world coordinate system is obtained. Combining the two equations above, we can obtain the linear relationship between the inverse projection of the pixel coordinate system to the 3D world coordinate system:

[0099]

[0100] To study the propagation of errors, we assume that there are certain errors in the estimation of the target's 3D world coordinates and pixel coordinates. Through linear transformation, we can obtain an approximate relationship for error propagation:

[0101]

[0102] In the formula, and These represent the errors in 3D world coordinates and pixel coordinates, respectively. For Jacobian matrices, Represented as:

[0103]

[0104] Assuming the detector's measurement error in the pixel coordinate system follows a Gaussian distribution, the measurement noise... It can be represented as:

[0105]

[0106] In the formula, Represents the noise factor. and These represent the width and height of the target, respectively.

[0107] Therefore, the covariance matrix of the measurement error in the three-dimensional world coordinate system is:

[0108]

[0109] Thus, the measurement noise is obtained by inversely projecting it from the pixel coordinate system to the three-dimensional world coordinate system.

[0110] Step 4: Extract feature points from the image background, and then apply the LK optical flow method to track the motion of these feature points between two frames. After obtaining the correspondence between the feature points in the two frames, calculate the affine transformation matrix of the image background motion using the RANSAC algorithm. Finally, apply the affine transformation matrix to the Kalman filter prediction results for camera motion compensation; specifically:

[0111] RANSAC calculation of affine transformation matrix:

[0112]

[0113] In the formula, It refers to the scaling and rotation components of an affine transformation. This is the translation part. The camera motion compensation for the predicted state and prediction covariance of the Kalman filter is calculated as follows:

[0114]

[0115]

[0116]

[0117] in, and for and The extended matrix; and These are the predicted state vectors of the Kalman filter before and after camera motion compensation, respectively. and These are the prediction covariance matrices of the Kalman filter before and after correction, respectively. After this, the update steps for the Kalman filter become:

[0118]

[0119] In the formula, For Kalman gain, The observation matrix; For updated state estimation, These are measured values; For the updated covariance estimate, It is an identity matrix.

[0120] Step 5: Similarity calculation method that integrates Mahalanobis distance and IoU. The method incorporates the measurement noise projected onto the ground plane into the Mahalanobis distance calculation, and adds IoU for auxiliary calculation; specifically:

[0121] Calculate the residuals:

[0122]

[0123] In the formula, For measured values, This is the observation matrix.

[0124] Calculate the residual covariance matrix:

[0125]

[0126] Therefore, the projected Mahalanobis distance can be calculated:

[0127]

[0128] use Impossible associations are excluded by calculating the threshold of the projected Mahalanobis distance for the 95% confidence interval. For a two-dimensional measurement space, the corresponding projected Mahalanobis distance threshold is... .

[0129] By combining projected Mahalanobis distance with IoU, a new similarity calculation method is obtained:

[0130]

[0131] in, Indicates the first The predicted value, the first One measurement value. Here, the weighting coefficients are used to balance the contributions of the projected Mahalanobis distance and IoU to the final cost, let... .

[0132] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions.

[0133] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0134] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0135] The above embodiments should be understood as illustrative only and not as limiting the scope of protection of the present invention. After reading the description of the present invention, those skilled in the art can make various alterations or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. A multi-target tracking method based on a linear camera model, characterized in that, Includes the following steps: Step 1: Obtain the input video sequence, perform target detection on each frame, and obtain the detection results of the target in the pixel coordinate system; Step 2: Perform lens distortion compensation on the detection results obtained in Step 1 to correct the imaging error caused by nonlinear distortion of the camera lens; Step 3: Based on the linear camera model, the detection results after lens distortion compensation are inversely projected from the pixel coordinate system to the ground plane in the three-dimensional world coordinate system to obtain the inverse projection related measurement parameters of the target; Step 4: After adding the relevant measurement parameters to the camera motion compensation, update the Kalman filter; Step 5: Based on the similarity calculation method of projection Mahalanobis distance and intersection-union ratio fusion, the similarity between the prediction result and the detection result is calculated to obtain the tracking trajectory.

2. The multi-target tracking method based on a linear camera model according to claim 1, characterized in that, Step 1 involves using a YOLOX detector to detect targets in the input video and obtaining detection results. YOLOX is a single-stage target detection neural network used to infer the target's category, confidence level, and bounding box coordinates from the input video frames.

3. The multi-target tracking method based on a linear camera model according to claim 1, characterized in that, Step 2: Lens distortion compensation includes radial distortion compensation and tangential distortion compensation; specifically: The mathematical model for radial distortion is: In the formula, The coordinates of the radial distortion point in the two-dimensional image plane coordinate system. The coordinates of the projection point on the two-dimensional image plane coordinate system. , , ... are the radial distortion coefficients corresponding to each order, respectively. This represents the distance between the ideal, distortion-free point and the principal point; Tangential distortion is caused by asymmetry in lens assembly. The mathematical model for tangential distortion is as follows: In the formula, The coordinates of the tangential distortion point in the two-dimensional image plane coordinate system are: , ... are the tangential distortion coefficients corresponding to each order; The distortion model of a camera lens is represented as follows: In the formula, These are the coordinates in the two-dimensional image plane coordinate system under the distortion model.

4. The multi-target tracking method based on a linear camera model according to claim 1, characterized in that, Step 3: Inversely project the pixel coordinate system onto the three-dimensional world coordinate system to obtain the distribution of relevant measurement parameters in the three-dimensional world coordinate system; specifically: According to the linear camera model, the transformation relationship between the world coordinate system and the pixel coordinate system is as follows: In the formula, As a scaling factor, In 3D camera coordinates A point on the axis; These are the coordinates in the pixel coordinate system. , ,in, Focal length and They represent direction and Focal length in direction, and They represent direction and Pixel density in the direction; Principal point coordinates are the pixel coordinates corresponding to the origin of the two-dimensional image plane coordinate system. Let be a rotation matrix. It is a translation vector; A point in a three-dimensional world coordinate system; It is an intrinsic matrix that contains the camera's internal parameters. This is the external matrix, which contains the camera's external parameters; Let be the projection matrix from the 3D world coordinate system to the pixel coordinate system, where is used for ease of subsequent calculations. ~ express matrix; make ,in, Assuming that the constant is constant and all target points lie in the same known depth plane, the above equation can be expressed as: make: In the formula, for ease of subsequent calculations, we use... ~ express matrix; Therefore, the projection matrix of the pixel coordinate system inversely projected onto the three-dimensional world coordinate system is obtained. Combining the two equations above, we can obtain the linear relationship between the inverse projection of the pixel coordinate system to the three-dimensional world coordinate system: Through linear transformation, an approximate relationship for error propagation is obtained: In the formula, and These represent the errors in 3D world coordinates and pixel coordinates, respectively. For Jacobian matrices, Represented as: ； Assuming the detector's measurement error in the pixel coordinate system follows a Gaussian distribution, the measurement noise... Represented as: ； In the formula, Indicates the noise factor. and These represent the width and height of the target, respectively. Therefore, the covariance matrix of the measurement error in the three-dimensional world coordinate system is: Thus, the measurement noise is obtained by inversely projecting it from the pixel coordinate system to the three-dimensional world coordinate system.

5. The multi-target tracking method based on a linear camera model according to claim 4, characterized in that, Step 4 specifically includes: extracting feature points from the image background, then applying the LK optical flow method to track the motion of the feature points between two frames; after obtaining the correspondence between the feature points in the two frames, calculating the affine transformation matrix of the image background motion using the RANSAC algorithm; finally, applying the affine transformation matrix to perform camera motion compensation on the Kalman filter prediction results; specifically: RANSAC calculation of affine transformation matrix: In the formula, It refers to the scaling and rotation components of an affine transformation. This is the translation part; the camera motion compensation for the predicted state and prediction covariance of the Kalman filter is calculated as follows: in, and for and The extended matrix; and These are the predicted state vectors of the Kalman filter before and after camera motion compensation, respectively. and These are the prediction covariance matrices of the Kalman filter before and after correction, respectively; after this, the update steps of the Kalman filter become: In the formula, For Kalman gain, The observation matrix; For the updated state estimate, These are measured values; For the updated covariance estimate, It is an identity matrix.

6. The multi-target tracking method based on a linear camera model according to claim 5, characterized in that, Step 5: The similarity calculation method that integrates Mahalanobis distance and IoU adds the measurement noise effect projected onto the ground plane to the Mahalanobis distance, and adds IoU for auxiliary calculation; specifically: Calculate the residuals: ； In the formula, For measured values, The observation matrix; Calculate the residual covariance matrix: ； Therefore, the projected Mahalanobis distance can be calculated: ； use Impossible associations are excluded by calculating the threshold of the projected Mahalanobis distance for the 95% confidence interval; for a two-dimensional measurement space, the corresponding projected Mahalanobis distance threshold is... ; By combining projected Mahalanobis distance with IoU, a new similarity calculation method is obtained: in, Indicates the first The predicted value, the first One measurement value; Here, the weighting coefficients are used to balance the contributions of the projected Mahalanobis distance and IoU to the final cost, let... .

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the multi-target tracking method based on a linear camera model as described in any one of claims 1 to 6.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the multi-target tracking method based on a linear camera model as described in any one of claims 1 to 6.

Images