A vehicle-mounted multi-camera target fusion method and system and storage medium

Abstract

The application discloses a kind of vehicle-mounted multi-camera target fusion method, system and storage medium, method includes: step 1: the image of vehicle-mounted multi-camera same time shooting, the intrinsic matrix of each camera, the extrinsic matrix of adjacent camera, the base distance between adjacent cameras are acquired, target detection is carried out using neural network, and all targets on each image are acquired;Step 2: determine non-overlapping target frame;Step 3: carry out uniform sampling to obtain sample point set, and according to the minimum criterion of polar line constraint error reproject to adjacent camera image, and obtain reprojected rectangular frame;Step 4: logical judgment is carried out to the original target frame of current camera, reprojected rectangular frame, adjacent camera target frame, and target fusion is carried out according to the judgment result;Step 5: all cameras are traversed, and steps 2-4 are repeated to process the target frame under all cameras.The application can efficiently and accurately judge the uniqueness of multi-camera target, and solve the multi-camera fusion problem of occluded target.

Description

Technical Field

[0001] This invention relates to the field of target fusion technology, and more specifically, to a method, system, and storage medium for target fusion among multiple vehicle-mounted cameras. Background Technology

[0002] Vision, as the core perception method for intelligent driving, can intuitively perceive dynamic and static targets during vehicle operation, ensuring safe driving. Due to limitations in the camera's field of view, multiple cameras are typically required to perceive a 360-degree view of the vehicle's surroundings, and images from adjacent cameras generally overlap. While ensuring full 360-degree perception, efficiently and accurately determining the uniqueness of targets in overlapping areas is also crucial for vehicle planning and control.

[0003] In the prior art, Chinese invention patent CN113077511B discloses a method and device for multi-camera target matching and tracking in automobiles. The method uses a projection matrix between the camera coordinate system and the vehicle coordinate system to project image targets onto the vehicle coordinate system. By calculating the difference in three-dimensional position coordinates and appearance similarity, it achieves the purpose of matching targets in multiple video frames. It is an industry consensus that monocular images lack depth information and cannot recover the full image depth information through a projection matrix for calculating the three-dimensional coordinates (x, y, z) in the world coordinate system. However, if the world coordinate z = 0 of a certain pixel is known, this pixel is generally called the ground point, and the x and y values ​​of that pixel in the world coordinate system can be calculated using the projection matrix. In this method, the vehicle coordinate system is considered the world coordinate system and located on a flat ground. The method of calculating coordinates in the vehicle coordinate system essentially assumes that the bottom edge of the target box is located on a horizontal plane, thus obtaining the x and y values ​​of the bottom edge ground point of the target box in the vehicle coordinate system. However, in actual driving, the road surface is usually not a standard level, and the bottom edge of the target bounding box is not necessarily on the ground where the vehicle is driving. For example, traffic signs may be in the sky, or sidewalks may be higher than the road surface. Therefore, the coordinates calculated using this method will have a large error. To improve accuracy, this method uses image similarity as a further basis for judgment. However, when target bounding boxes are mutually occluded, using image similarity is not only time-consuming to calculate, but also cannot accurately distinguish occluded targets.

[0004] Therefore, the above methods have certain limitations. A more efficient and accurate multi-camera target fusion method is needed that is applicable to different road surfaces, does not restrict the position of the target in the image, and can solve the problem of partial occlusion of the target during vehicle operation. Summary of the Invention

[0005] To address the problems of the prior art, this invention provides a method, system, and storage medium for target fusion among multiple vehicle cameras. This method can efficiently and accurately determine the uniqueness of targets among multiple cameras without limiting vehicle driving conditions, while also solving the problem of fusion among multiple cameras for occluded targets.

[0006] The primary objective of this invention is to solve the aforementioned technical problems. The technical solution of this invention is as follows:

[0007] The first aspect of this invention provides a target fusion method among multiple on-vehicle cameras, comprising the following steps:

[0008] Step 1: Obtain images captured simultaneously by multiple onboard cameras, the intrinsic parameter matrix of each camera, the extrinsic parameter matrix of adjacent cameras, and the base distance between adjacent cameras. Use a neural network to perform target detection and obtain all targets in each image.

[0009] Step 2: Select each camera to determine the non-overlapping bounding box;

[0010] Step 3: Uniformly sample the region within the non-overlapping target box to obtain a set of sampling points. Reproject the set of sampling points onto the adjacent camera image according to the epipolar constraint error minimization criterion to obtain a reprojected rectangle.

[0011] Step 4: Perform logical judgments on the original target bounding box, reprojected rectangle, and target bounding boxes of the current camera, and perform target fusion processing based on the judgment results;

[0012] Step 5: Traverse all cameras and repeat steps 2-4 to process the bounding boxes under all cameras.

[0013] Furthermore, the neural network includes: convolutional neural network and transform network.

[0014] Furthermore, for each camera, non-overlapping bounding boxes are determined. The specific steps are as follows:

[0015] Determine whether there is overlap between the original target boxes under the currently selected camera. If there is overlap, obtain the non-overlapping area of ​​each target box as the non-overlapping target box; if there is no overlap, use the target box itself as the non-overlapping target box.

[0016] Furthermore, the set of sampling points is reprojected onto adjacent camera images according to the epipolar constraint error minimization criterion to obtain reprojected bounding boxes. The specific steps are as follows:

[0017] (a) The distortion-free coordinates Po are obtained by performing distortion correction on the sampling points;

[0018] (b) Use the intrinsic parameter matrix of each camera and the extrinsic parameter matrix of the adjacent cameras to obtain the right eye's intermediate target-based fundamental matrix F and the distortion-free coordinates Po of the sampling point, and calculate the epipolar equation corresponding to the sampling point Po in the adjacent camera image.

[0019] le = FPo

[0020] (c) Based on the epipolar equation le, set a distance threshold Td, and according to the formula for calculating the direct distance between a point and the epipolar line, obtain the distance Pt of all pixels on the image whose distance from the epipolar line is less than Td. i (t i =0,1,...,n);

[0021] (d) Based on sampling point Po and pixel point Pt i For the pixel Pt i Perform 3D reconstruction to obtain sampling point Po and pixel point Pt. i The three-dimensional coordinates Wt(Xt,Yt,Zt) of two corresponding points in the coordinate system of adjacent cameras;

[0022] (e) Based on the extrinsic parameter matrices R and T of the current camera and the adjacent camera, convert the three-dimensional coordinates in the coordinate system of the adjacent camera to the three-dimensional coordinates Wo in the coordinate system of the current camera;

[0023] (f) Based on the intrinsic parameter matrix of the current camera, convert the three-dimensional coordinates of the sampling points in the current camera coordinate system to the reprojected image coordinates in the current camera coordinate system. The converted points are denoted as Pr. i (i = 0, 1, ..., n);

[0024] (g) Calculate the distance between the sampling point Po and the transformed point to obtain the reprojected pixel coordinates with the smallest distance in Pr. i The index in the table determines the optimal reprojection point Pt;

[0025] (h) Determine the largest bounding rectangle based on the best reprojection point of all sampling points, which is used as the reprojection rectangle of the current camera's unobstructed rectangle on the adjacent camera image.

[0026] Furthermore, the expression for 3D reconstruction is:

[0027] Zt=B*f / d

[0028] Xt=Z*Pt i _x / f

[0029] Yt=Z*Pt i _y / f

[0030] Where B is the base distance between the current camera and the adjacent camera, f is the camera focal length, and d is the distance between Po and Pt. i Parallax of a point, Pti _x is point Pt i The pixel x-coordinate, Pt i _y represents point Pt i The pixel y-coordinate.

[0031] Furthermore, the three-dimensional coordinate transformation formula is as follows:

[0032] Wo=RWt+T

[0033] Wo represents the three-dimensional coordinates in the current camera coordinate system.

[0034] Furthermore, step 4 involves performing logical judgments on the original target bounding box, reprojected rectangle, and target bounding boxes of the current camera, and then performing target fusion processing based on the judgment results. Specifically:

[0035] Calculate the intersection-union ratio (CIU) between the reprojected rectangle and the adjacent camera target boxes, and set the CIU threshold to Tiou;

[0036] If the intersection-union ratio (IU) of adjacent camera target boxes and reprojected rectangles is greater than the threshold Tiou, then the original target boxes of the current camera corresponding to the adjacent camera target boxes and reprojected boxes are determined to be one target, and target fusion is performed.

[0037] If the intersection-union ratio (IU) of adjacent camera target boxes and reprojected rectangles is less than or equal to the threshold Tiou, then the original target boxes of the current camera corresponding to the adjacent camera target boxes and the reprojected boxes are not the same target, and target fusion is not performed.

[0038] A second aspect of the present invention provides a vehicle-mounted multi-camera target fusion system, the system comprising: a memory and a processor, wherein the memory includes a program for a vehicle-mounted multi-camera target fusion method, and when the processor executes the program for the vehicle-mounted multi-camera target fusion method, the program implements the steps of the vehicle-mounted multi-camera target fusion method as follows:

[0039] Step 1: Obtain images captured simultaneously by multiple onboard cameras, the intrinsic parameter matrix of each camera, the extrinsic parameter matrix of adjacent cameras, and the base distance between adjacent cameras. Use a neural network to perform target detection and obtain all targets in each image.

[0040] Step 2: Select each camera to determine the non-overlapping bounding box;

[0041] Step 3: Uniformly sample the region within the non-overlapping target box to obtain a set of sampling points. Reproject the set of sampling points onto the adjacent camera image according to the epipolar constraint error minimization criterion to obtain a reprojection rectangle.

[0042] Step 4: Perform logical judgments on the original target bounding box, reprojected rectangle, and target bounding boxes of the current camera, and perform target fusion processing based on the judgment results;

[0043] Step 5: Traverse all cameras and repeat steps 2-4 to process the bounding boxes under all cameras.

[0044] Furthermore, for each camera, non-overlapping bounding boxes are determined. The specific steps are as follows:

[0045] Determine whether there is overlap between the original target boxes under the currently selected camera. If there is overlap, obtain the non-overlapping area of ​​each target box as the non-overlapping target box; if there is no overlap, use the target box itself as the non-overlapping target box.

[0046] A third aspect of the present invention provides a storage medium including a target fusion method program between vehicle-mounted multi-cameras, wherein when the target fusion method program between vehicle-mounted multi-cameras is executed by a processor, the steps of the aforementioned target fusion method between vehicle-mounted multi-cameras are implemented.

[0047] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0048] This invention uses the epipolar constraint error minimization criterion to obtain the reprojected rectangular target box from the sampling point set, which can efficiently and accurately determine the uniqueness of targets among multiple cameras without limiting vehicle driving conditions; at the same time, by uniformly sampling the region within the non-overlapping target box, it solves the problem of fusion between multiple cameras for occluded targets. Attached Figure Description

[0049] Figure 1 This is a flowchart of a target fusion method between multiple cameras in a vehicle according to an embodiment of the present invention.

[0050] Figure 2 A flowchart for determining non-overlapping target boxes in an embodiment of the present invention. Detailed Implementation

[0051] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0052] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0053] Example 1

[0054] like Figure 1 As shown, the first aspect of the present invention provides a target fusion method among multiple vehicle cameras, comprising the following steps:

[0055] Step 1: Obtain images captured simultaneously by multiple onboard cameras, the intrinsic parameter matrix of each camera, the extrinsic parameter matrix of adjacent cameras, and the base distance between adjacent cameras. Use a neural network to perform target detection and obtain all targets in each image.

[0056] It should be noted that in this invention, the intrinsic parameter matrix describes the relationship between the three-dimensional world coordinates and the two-dimensional pixel coordinate system. A definite image pixel coordinate can be obtained from the three-dimensional coordinates through the intrinsic parameter matrix, and a pixel coordinate can be obtained from multiple three-dimensional world coordinates (not unique solutions) through the intrinsic parameter matrix; these are parameters for each camera. The extrinsic parameter matrix transforms three-dimensional coordinate points from one coordinate system to another, describing the pose transformation between different cameras. Both the intrinsic and extrinsic parameter matrices can be obtained through camera calibration methods, which are numerous and universal; this invention does not limit the methods used.

[0057] This invention utilizes neural networks for object detection. Each detected object in an image corresponds to a bounding box, and the object is selected by the bounding box. This invention is not limited to a specific neural network; it can use convolutional neural networks such as one-stage object detection networks (e.g., the YOLO series), or two-stage object detection networks such as Faster R-CNN and its optimized versions. Transform networks, such as DETR and its variants, can also be used.

[0058] Step 2: Select each camera to determine the non-overlapping bounding box;

[0059] Figure 2 The process for determining non-overlapping target boxes is shown.

[0060] It should be noted that the specific steps for selecting each camera to determine the non-overlapping bounding box are as follows:

[0061] Determine whether there is overlap between the original target boxes under the currently selected camera. If there is overlap, obtain the non-overlapping area of ​​each target box as the non-overlapping target box.

[0062] If there is no overlap, the target box itself is used as the non-overlapping target box.

[0063] It should be noted that by determining the non-overlapping target bounding box and then the non-overlapping region, it is easier to sample the non-overlapping region to obtain a set of sampling points.

[0064] Step 3: Uniformly sample the region within the non-overlapping target box to obtain a set of sampling points. Reproject the set of sampling points onto the adjacent camera image according to the epipolar constraint error minimization criterion to obtain a reprojected rectangle.

[0065] It should be noted that the epipolar constraint error minimization criterion can accurately establish a target box in one camera image coordinate system and a reprojected rectangle in another camera image coordinate system.

[0066] The set of sampling points is reprojected onto adjacent camera images according to the epipolar constraint error minimization criterion to obtain reprojected bounding boxes. The specific steps are as follows:

[0067] (a) The distortion-free coordinates Po are obtained by performing distortion correction on the sampling points;

[0068] (b) Use the intrinsic parameter matrix of each camera and the extrinsic parameter matrix of the adjacent cameras to obtain the right eye's intermediate target-based fundamental matrix F and the distortion-free coordinates Po of the sampling point, and calculate the epipolar equation corresponding to the sampling point Po in the adjacent camera image.

[0069] le = FPo

[0070] (c) Based on the epipolar equation le, set a distance threshold Td, and according to the formula for calculating the direct distance between a point and the epipolar line, obtain the distance Pt of all pixels on the image whose distance from the epipolar line is less than Td. i (t i =0,1,...,n);

[0071] (d) Based on sampling point Po and pixel point Pt i For the pixel Pt i Perform 3D reconstruction to obtain sampling point Po and pixel point Pt. i The three-dimensional coordinates Wt(Xt,Yt,Zt) of two corresponding points in the coordinate system of adjacent cameras;

[0072] The expression for 3D reconstruction is:

[0073] Zt=B*f / d

[0074] Xt=Z*Pt i _x / f

[0075] Yt=Z*Pt i _y / f

[0076] Where B is the base distance between the current camera and the adjacent camera, f is the camera focal length, and d is the distance between Po and Pt. i Parallax of a point, Pt i _x is point Pt i The pixel x-coordinate, Pt i _y represents point Pt i The pixel y-coordinate.

[0077] (e) Based on the extrinsic parameter matrices R and T of the current camera and the adjacent cameras, convert the 3D coordinates in the coordinate system of the adjacent cameras to the 3D coordinates Wo in the coordinate system of the current camera;

[0078] The formula for 3D coordinate transformation is:

[0079] Wo=RWt+T

[0080] Wo represents the three-dimensional coordinates in the current camera coordinate system.

[0081] (f) Based on the intrinsic parameter matrix of the current camera, transform the 3D coordinates of the sampling points in the current camera coordinate system into the reprojected image coordinates in the current camera coordinate system. The transformed points are denoted as Pr. i (i = 0, 1, ..., n);

[0082] (g) Calculate the distance between the sampling point Po and the transformed point to obtain the reprojected pixel coordinates with the smallest distance in Pr. i The index in the table determines the optimal reprojection point Pt;

[0083] (h) Determine the largest bounding rectangle based on the best reprojection point of all sampling points, which is used as the reprojection rectangle of the current camera's unobstructed rectangle on the adjacent camera image.

[0084] Step 4: Perform logical judgments on the original target bounding box, reprojected rectangle, and target bounding boxes of the current camera, and perform target fusion processing based on the judgment results;

[0085] It should be noted that the target fusion process involves logically judging the original target bounding box, reprojected rectangle, and target bounding boxes of the current camera, and then performing target fusion based on the judgment results. The specific process is as follows:

[0086] Calculate the intersection-union ratio (CIU) between the reprojected rectangle and the adjacent camera target boxes, and set the CIU threshold to Tiou;

[0087] If the intersection-union ratio (IU) of adjacent camera target boxes and reprojected rectangles is greater than the threshold Tiou, then the original target boxes of the current camera corresponding to the adjacent camera target boxes and reprojected boxes are determined to be one target, and target fusion is performed.

[0088] If the intersection-union ratio (IU) of adjacent camera target boxes and reprojected rectangles is less than or equal to the threshold Tiou, then the original target boxes of the current camera corresponding to the adjacent camera target boxes and the reprojected boxes are not the same target, and target fusion is not performed.

[0089] Step 5: Traverse all cameras and repeat steps 2-4 to process the bounding boxes under all cameras.

[0090] A second aspect of the present invention provides a vehicle-mounted multi-camera target fusion system, the system comprising: a memory and a processor, wherein the memory includes a program for a vehicle-mounted multi-camera target fusion method, and when the processor executes the program for the vehicle-mounted multi-camera target fusion method, the program implements the steps of the vehicle-mounted multi-camera target fusion method as follows:

[0091] Step 1: Obtain images captured simultaneously by multiple onboard cameras, the intrinsic parameter matrix of each camera, the extrinsic parameter matrix of adjacent cameras, and the base distance between adjacent cameras. Use a neural network to perform target detection and obtain all targets in each image.

[0092] Step 2: Select each camera to determine the non-overlapping bounding box;

[0093] Step 3: Uniformly sample the region within the non-overlapping target box to obtain a set of sampling points. Reproject the set of sampling points onto the adjacent camera image according to the epipolar constraint error minimization criterion to obtain a reprojected rectangle.

[0094] Step 4: Perform logical judgments on the original target bounding box, reprojected rectangle, and target bounding boxes of the current camera, and perform target fusion processing based on the judgment results;

[0095] Step 5: Traverse all cameras and repeat steps 2-4 to process the bounding boxes under all cameras.

[0096] Furthermore, for each camera, non-overlapping bounding boxes are determined. The specific steps are as follows:

[0097] Determine whether there is overlap between the original target boxes under the currently selected camera. If there is overlap, obtain the non-overlapping area of ​​each target box as the non-overlapping target box; if there is no overlap, use the target box itself as the non-overlapping target box.

[0098] A third aspect of the present invention provides a storage medium including a target fusion method program between vehicle-mounted multi-cameras, wherein when the target fusion method program between vehicle-mounted multi-cameras is executed by a processor, the steps of the aforementioned target fusion method between vehicle-mounted multi-cameras are implemented.

[0099] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for fusing a target between multiple cameras on a vehicle, the method comprising: Includes the following steps: Step 1: Obtain images captured simultaneously by multiple onboard cameras, the intrinsic parameter matrix of each camera, the extrinsic parameter matrix of adjacent cameras, and the base distance between adjacent cameras. Use a neural network to perform target detection and obtain all targets in each image. Step 2: Select each camera to determine the non-overlapping bounding box; Step 3: Uniformly sample the region within the non-overlapping target box to obtain a set of sampling points. Reproject this set of sampling points onto adjacent camera images according to the epipolar constraint error minimization criterion to obtain a reprojected rectangle. The specific steps are as follows: (a) The distortion-free coordinates Po are obtained by performing distortion correction on the sampling points; (b) Use the intrinsic parameter matrix of each camera and the extrinsic parameter matrix of the adjacent cameras to obtain the right eye's basis matrix F based on the intermediate object and the distortion-free coordinates Po of the sampling point, and calculate the epipolar equation corresponding to the sampling point Po in the image of the adjacent camera. le = FPo (c) According to the epipolar equation le, set the distance threshold Td, and according to the point-to-direct distance calculation formula, obtain all pixel points Pt on the image with a distance to the epipolar line less than Td i (t i =0,1,...,n) (d) based on the sampling point Po and the pixel point Pt i , the pixel point Pt i performing three-dimensional reconstruction to obtain the sampling point Po and the pixel point Pt i three-dimensional coordinates Wt(Xt, Yt, Zt) of the two corresponding points in the adjacent camera coordinate system (e) Based on the extrinsic parameter matrices R and T of the current camera and the adjacent cameras, convert the three-dimensional coordinates in the coordinate system of the adjacent cameras to the three-dimensional coordinates Wo in the coordinate system of the current camera; (f) According to the intrinsic matrix of the current camera, the three-dimensional coordinates of the sampling points in the current camera coordinate system are converted into the re-projection image coordinates under the current camera, and the converted points are denoted as Pr i (i = 0, 1,..., n); (g) calculate the distance between the sampling point Po and the converted point, get the index of the re-projection pixel coordinate with the minimum distance in Pr i , determine the best re-projection point Pt; (h) Determine the largest bounding rectangle based on the best reprojection point of all sampling points, and use it as the reprojection rectangle of the current camera's unobstructed rectangle on the adjacent camera image. Step 4: Perform logical judgments on the original target bounding box, reprojected rectangle, and target bounding boxes of the current camera, and perform target fusion processing based on the judgment results; Step 5: Traverse all cameras and repeat steps 2-4 to process the bounding boxes under all cameras.

2. The method according to claim 1, wherein, The neural networks include: convolutional neural networks and transform networks.

3. The method of claim 1, wherein, The specific steps for selecting each camera and determining the non-overlapping bounding box are as follows: Determine whether there is overlap between the original target boxes under the currently selected camera. If there is overlap, obtain the non-overlapping area of ​​each target box as the non-overlapping target box; if there is no overlap, use the target box itself as the non-overlapping target box.

4. The target fusion method among multiple vehicle-mounted cameras according to claim 1, characterized in that, The expression for 3D reconstruction is: Zt = B*f / d Xt = Z*Pt i _x / f Yt = Z*Pt i _y / f where B is the base distance between the current camera and the neighboring camera, f is the camera focal length, d is the parallax between the Po and Pt i points i _x is the pixel x coordinate of the point Pt i . i _y is the pixel y coordinate of the point Pt i .

5. The target fusion method among multiple vehicle-mounted cameras according to claim 1, characterized in that, The formula for 3D coordinate transformation is: Wo = RWt+T Wo represents the three-dimensional coordinates in the current camera coordinate system.

6. The target fusion method among multiple vehicle-mounted cameras according to claim 1, characterized in that, Step 4 involves performing logical judgments on the original target bounding box, reprojected rectangle, and target bounding boxes of the current camera, and then performing target fusion processing based on the judgment results. Specifically: Calculate the intersection-union ratio (CIU) between the reprojected rectangle and the adjacent camera target boxes, and set the CIU threshold to Tiou; If the intersection-union ratio (IU) of adjacent camera target boxes and reprojected rectangles is greater than the threshold Tiou, then the original target boxes of the current camera corresponding to the adjacent camera target boxes and reprojected boxes are determined to be one target, and target fusion is performed. If the intersection-union ratio (IU) of adjacent camera target boxes and reprojected rectangles is less than or equal to the threshold Tiou, then the original target boxes of the current camera corresponding to the adjacent camera target boxes and the reprojected boxes are not the same target, and target fusion is not performed.

7. A vehicle-mounted multi-camera target fusion system, characterized in that, The system includes: a memory and a processor. The memory includes a program for a target fusion method between vehicle-mounted multi-cameras. When executed by the processor, the program for a target fusion method between vehicle-mounted multi-cameras implements the steps of the target fusion method between vehicle-mounted multi-cameras as described in any one of claims 1-6. Step 1: Obtain images captured simultaneously by multiple onboard cameras, the intrinsic parameter matrix of each camera, the extrinsic parameter matrix of adjacent cameras, and the base distance between adjacent cameras. Use a neural network to perform target detection and obtain all targets in each image. Step 2: Select each camera to determine the non-overlapping bounding box; Step 3: Uniformly sample the region within the non-overlapping target box to obtain a set of sampling points. Reproject the set of sampling points onto the adjacent camera image according to the epipolar constraint error minimization criterion to obtain a reprojected rectangle. Step 4: Perform logical judgments on the original target bounding box, reprojected rectangle, and target bounding boxes of the current camera, and perform target fusion processing based on the judgment results; Step 5: Traverse all cameras and repeat steps 2-4 to process the bounding boxes under all cameras.

8. The vehicle-mounted multi-camera target fusion system according to claim 7, characterized in that, The specific steps for selecting each camera and determining the non-overlapping bounding box are as follows: Determine whether there is overlap between the original target boxes under the currently selected camera. If there is overlap, obtain the non-overlapping area of ​​each target box as the non-overlapping target box; if there is no overlap, use the target box itself as the non-overlapping target box.

9. A storage medium, characterized in that, The storage medium includes a target fusion method program between vehicle-mounted multi-cameras. When the target fusion method program between vehicle-mounted multi-cameras is executed by a processor, it implements the steps of a target fusion method between vehicle-mounted multi-cameras as described in any one of claims 1 to 6.

Images