Motion multi-target matching and position estimation method based on asynchronous image sequence

Abstract

The invention discloses a motion multi-target matching and position estimation method based on an asynchronous image sequence, and belongs to the fields of computer vision, SLAM, photogrammetry and the like. Firstly, three cameras are adopted to shoot sequence images, three space planes are established for each target, candidate tracks of the targets are determined for every two planes, candidate positions at specific heights are calculated, and then a candidate position matrix of multiple targets between the camera 1 and the camera 2 and a candidate position matrix of multiple targets between the camera 1 and the camera 3 are obtained; secondly, establishing a flexible search method according to a principle that the distance is minimum as a correct matching principle, performing undifferentiated distance calculation on the two matrixes to obtain multi-target matching under the condition of the minimum distance, and establishing a position estimation algorithm based on line-plane constraint according to a matching result; according to the method, the problem that the existing algorithm cannot perform multi-target matching and position estimation under the condition of asynchronism and no obvious characteristics is solved, no relation exists between cameras in the calculation process, the adaptability is higher, and the equipment structure is simpler.

Description

Technical Field

[0001] The invention belongs to the fields of computer vision, SLAM, photogrammetry, etc., and specifically relates to a moving multi-target matching and position estimation method based on asynchronous image sequences. Background Art

[0002] In computer vision, SLAM, photogrammetry and other related fields, feature matching technology is a core component, which involves extracting feature points from multiple images and matching these feature points to determine the correspondence between images. This technology has a wide range of applications in many fields, including but not limited to image stitching, 3D reconstruction, target recognition, visual tracking, etc.

[0003] In technical details, image feature matching usually includes three basic steps: feature point detection, feature description, and feature matching. Feature point detection aims to identify significant structural features from the image, such as corners, edges, or significant morphological regions. These feature points are usually represented as coordinate points in the image, which should remain stable under different viewing angles or lighting conditions. Feature description involves generating descriptors for these feature points. The descriptor is a quantitative representation of the neighborhood around the feature point, which is used to capture the local information of the feature point. Finally, the feature matching process finds matching feature point pairs by comparing the descriptors of the feature points in the two images. With the development of deep learning technology, the field of image feature matching has also ushered in new development opportunities. Deep learning methods can learn pixel-level matching relationships directly from image pairs, including learning more accurate feature point sets, learning the main directions or scales of feature points, and learning feature descriptors with discriminative and matchable capabilities. For example, SuperPoint [Superpoint: Self-supervised interest point detection and description] is a self-supervised training method that is trained on artificially generated data and can then obtain results on real natural images. D2-Net[D2-net:A trainable cnn for joint description and detection of local features] uses deep information to train the network so that the extracted features can be correctly matched.

[0004] For the above conventional methods, the features in the image must be obvious, and the image must be still or multiple images must be synchronized in time to obtain the features with the same name. However, in some scenes, multiple targets fall from the air, and there is no synchronization device between multiple cameras, so that the captured images are neither synchronized nor have obvious features. In this case, the above methods cannot match multiple moving targets, let alone estimate the positions of multiple targets. Summary of the invention

[0005] (I) Purpose of the invention

[0006] The purpose of the present invention is to provide a method for motion multi-target matching and position estimation based on asynchronous image sequences, which uses asynchronous image sequences to perform multi-target matching and position estimation of linear motion in dynamic scenes, and solves the problem that existing algorithms cannot perform multi-target matching and position estimation in asynchronous and non-obvious features. In the present invention, each camera establishes a plane constraint for each target through an image sequence, so three cameras can establish three surface constraints for each target. The surface constraints of the two cameras intersect indiscriminately and are solved at a specific height. Two distance candidate matrices can be established, and the final matching result is obtained by minimizing the position distance. Finally, line-surface constraints are used for position estimation.

[0007] (II) Technical solution

[0008] In order to achieve the above objectives and solve the above technical problems, the technical solutions of the present invention are as follows:

[0009] The method for matching and estimating the position of multiple moving targets based on an asynchronous image sequence specifically includes the following steps:

[0010] Step 1: Single target position estimation based on asynchronous image sequences

[0011] Assuming the target moves in a straight line, the two cameras capture asynchronous images of the target;

[0012] Assume that the left camera and the right camera capture multiple frames of target images respectively, and both the left camera and the right camera are calibrated. The conversion relationship between the camera coordinate system Oc1_Xc1Yc1Zc1, Oc2_Xc2Yc2Zc2 and the world coordinate system Ow_XwYwZw is known, satisfying the following relationship:

[0013]

[0014] The above formula is expressed as a 3D point (X w Y w Z w ) is projected on the left camera as (u c1 v c1 );

[0015] Among them, Z c1 is the scale factor, unknown; M c1 is the projection matrix of the left camera, which is obtained through prior calibration and is known;

[0016] The left camera captures images of multiple targets, and a straight line is fitted on the image plane as follows:

[0017]

[0018] Among them (a c1 b c1 ) is the coefficient of the fitted straight line;

[0019] Then the expression of the left camera plane where the target is located is as follows

[0020]

[0021] The expression of the right camera plane where the target is located based on the multi-frame target imaging position of the right camera is as follows:

[0022]

[0023] Among them, M c2 is the projection matrix of the right camera, obtained through prior calibration and known; (a c2 b c2 ) is the coefficient of the fitting line on the right camera image plane, and the target trajectory is the intersection of the left and right planes; the spatial position is used as the target matching information, and the target position is the intersection of a straight line and a plane. The calculation formula is as follows:

[0024]

[0025] In this way, the position estimation of a single target is achieved, and h is the height of the target;

[0026] Step 2: Multi-target matching and position estimation at a specific height

[0027] Assume that there are n targets entering the field of view, and use multiple frames of images to fit n straight lines on the image plane of camera 1. The expression of the nth straight line is as follows:

[0028]

[0029] Similarly, n straight lines are fitted on the image plane of camera 2 The expression of the nth straight line is as follows:

[0030]

[0031] Fitting the two sets of straight lines Treat it as a blind box and perform indiscriminate matching, that is, the i-th fitting line l in camera 1 i and all fitted lines in camera 2 Match them once respectively, 1≤i≤n, and then use the single target position estimation method in step 1 to obtain the candidate value of the i-th target position in camera 1 after each match:

[0032]

[0033] here represents the candidate value after the i-th fitting line in camera 1 and all the fitting lines in camera 2 are matched one by one, represents the target position after the i-th fitting line in camera 1 and the j-th fitting line in camera 2 are matched; perform the above operation on each fitting line in camera 1 to obtain a distance candidate matrix as follows:

[0034]

[0035] Introduce the third camera, and then similarly obtain a distance candidate matrix for camera 1 and camera 3, as follows:

[0036]

[0037] Under correct matching, the i-th target in camera 1 corresponds to two sets of candidate values, namely

[0038]

[0039] In these two sets of data, only one of each set is the result of correct matching, and under the condition of no error, the data set There is a value and data set The values ​​in are exactly equal, and the equal values ​​are the correct match. The situation is considered a correct match;

[0040] Update the matrix as follows

[0041]

[0042] Matching target is The distance between the two values ​​is the smallest (due to the existence of errors, it may be the second smallest), and n groups of such values ​​are found to minimize the sum of the distances. Each such pairing is the optimal match. After updating the matrix, repeat the above steps to obtain the correct matches of all other targets.

[0043] Step 3: Position estimation at a specific moment

[0044] The position of the target in camera 2 at each moment is measured by line-surface intersection. Now the matching is completed, so here only cameras 1 and 2 are used to estimate the position of a single target. At time t, camera 2 captures the target and forms an image at x. t (u t v t ); Camera 1 did not capture the target at this moment, but captured the target at other moments; According to Formula 3, a plane S is obtained, according to the imaging x t , it can be determined that the target is located on the straight line l at time t t The expression of the straight line is as follows:

[0045]

[0046] Where a is an arbitrary scale factor, R c2 ,t c2 are the rotation matrix and translation vector of camera c2 respectively, which are obtained by PnP algorithm. Combining formula (3) and (14), we can get the position of the target at time t, which is the straight line l t The intersection point with plane S.

[0047] (III) Effective income

[0048] 1. The present invention proposes a method for matching and estimating the positions of multiple moving targets using asynchronous image sequences, which solves the problem of matching and estimating the positions of multiple targets in featureless and asynchronous situations.

[0049] 2. The present invention establishes a candidate position matrix through face-face constraints and obtains the final matching result by minimizing the position distance. Finally, the position is estimated using line-face constraints. Compared with the prior art, the present invention does not require rich features in the field of view, does not require epipolar constraints after camera synchronization, has no relationship between cameras, and makes full use of multiple frames of images taken.

[0050] 3. The present invention has stronger adaptability and a simpler device structure because synchronization is not required. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 Schematic diagram of the imaging geometry of two cameras with asynchronous image sequences;

[0052] Figure 2 Schematic diagram of the geometric structure of single target position estimation;

[0053] Figure 3 Schematic diagram of multi-target imaging;

[0054] Figure 4 Schematic diagram of the flexible optimal process;

[0055] Figure 5 Distance and schematic diagram at each matching step;

[0056] Figure 6 Schematic diagram of target position calculation at a specific moment;

[0057] Figure 7 Schematic diagram of an implementation scenario of Embodiment 1 of the present invention;

[0058] Figure 8 Schematic diagram of trajectory simulation according to an embodiment of the present invention;

[0059] Fig. 9 Schematic diagram of real images captured by three cameras in an embodiment of the present invention;

[0060] Fig.10 A schematic diagram of the first target matching of an embodiment of the present invention;

[0061] Fig.11 Schematic diagram of updating the matrix after the first target matching in an embodiment of the present invention. DETAILED DESCRIPTION

[0062] The present invention is explained and illustrated in detail below in conjunction with the embodiments and drawings.

[0063] In order to complete the positioning of multiple moving targets, the targets need to be matched. Here, due to the single background and very few features, such as the sky, the features cannot be extracted, and thus the matching cannot be performed by the traditional feature matching algorithm. In addition, due to the continuous movement of the target, the acquired images are asynchronous, so that the matching algorithm based on the epipolar constraint cannot complete the matching. Due to the inability to complete the matching and the asynchronous images, the traditional stereoscopic vision measurement algorithm cannot be used to estimate the position of the moving multiple targets. To solve the above problems, the present invention proposes a new algorithm for matching and position estimation of moving multiple targets based on asynchronous image sequences. First, three cameras are used to shoot sequence images, three spatial planes are established for each target, candidate trajectories of the target are determined for every two planes, and candidate positions at a specific height are calculated, and then the candidate position matrix of the multiple targets between camera 1 and camera 2 and the candidate position matrix of the multiple targets between camera 1 and camera 3 are obtained; then, according to the principle of minimum distance as the correct match, a flexible search method is established, and the two matrices are calculated by indifferent distance to obtain the multi-target matching under the minimum distance, and according to the matching results, a position estimation algorithm based on line and surface constraints is established.

[0064] The present invention proposes a method for matching and estimating the positions of multiple moving objects based on asynchronous image sequences, which can match linear moving objects captured by multiple cameras and then estimate the positions of multiple objects. To achieve the above purpose, the present invention adopts the following technical solutions:

[0065] Step 1: Single target position estimation based on asynchronous image sequences

[0066] Using multiple frames in the image sequence, a position estimation method under surface constraints is constructed. Here, it is assumed that the target moves in a straight line and the two cameras capture asynchronous images of the target. Figure 1 shown.

[0067] Assume that the left camera and the right camera capture multiple frames of target images respectively, such as Figure 1 As shown. In the figure, both the left and right cameras are calibrated, that is, the conversion relationship between the camera coordinate system Oc1_Xc1Yc1Zc1, Oc2_Xc2Yc2Zc2 and the world coordinate system Ow_XwYwZw is known, satisfying the following relationship

[0068]

[0069] The above formula is expressed as a 3D point (X w Y w Z w ) is projected on the left camera as (u c1 v c1 ). Among them, Z c1 is the scale factor, unknown; M c1 is the projection matrix of the left camera, which is obtained through prior calibration and is known. The left camera captures images of multiple targets, so a straight line can be fitted on the image plane as follows

[0070]

[0071] Among them (a c1 b c1 ) is the coefficient of the fitted line. Combining equations (1)-(2), we can eliminate The expression of the left camera plane where the target is located can be obtained as follows

[0072]

[0073] Similarly, according to the multi-frame target imaging position of the right camera, the expression of the right camera plane where the target is located is obtained as follows.

[0074]

[0075] Among them, M c2 is the projection matrix of the right camera, obtained through prior calibration and known; (a c2 b c2 ) is the coefficient of the fitted straight line on the right camera image plane. Combining equations 3-4, we can get the trajectory of the target, which is the intersection of the left and right planes. The spatial straight line is a complex expression. In order to facilitate multi-target matching later, we use the spatial position as the matching information. Here, the spatial position of the target at height h is calculated as follows: Figure 2 shown.

[0076] It can be seen that the target position here is the intersection of a straight line (trajectory) and a plane, and the calculation formula is as follows.

[0077]

[0078] In this way, the position estimation of a single target is achieved.

[0079] Step 2: Multi-target matching and position estimation at a specific height

[0080] Assume that multiple moving targets appear in the field of view of three cameras at the same time, and their images in camera 1 are as follows: Figure 3 shown.

[0081] like Figure 3 , assuming that there are n targets entering the field of view, using multiple frames of images, n straight lines can be fitted on the image plane of camera 1 The expression of the nth straight line is as follows.

[0082]

[0083] Similarly, n straight lines can be fitted on the image plane of camera 2 The expression of the nth straight line is as follows.

[0084]

[0085] Here, in the case of multiple targets, we do not know the straight line obtained by fitting and Therefore, the target position cannot be directly obtained like the single target position estimation. Here, the two sets of fitting lines are regarded as blind boxes and matched indiscriminately, that is, the i-th fitting line l in camera 1 i (1≤i≤n) and all fitted lines in camera 2 Then, the single target position estimation method is used to obtain the possible candidate values ​​of the i-th target position in camera 1 after each match.

[0086]

[0087] here represents the candidate value after the i-th fitting line (i-th target) in camera 1 and all the fitting lines in camera 2 are matched one by one, It represents the target position after the i-th fitting line (i-th target) in camera 1 and the j-th fitting line in camera 2 are matched. Then, we perform the above operation on each fitting line in camera 1 to obtain a distance candidate matrix as follows.

[0088]

[0089] It can be seen that we cannot determine the specific matching results, resulting in the change from the correct n matches to n×n matches.

[0090] In order to determine the correct multi-target matching in an asynchronous image sequence, we introduce a third camera, and then obtain a distance candidate matrix for camera 1 and camera 3 in the same way, as follows.

[0091]

[0092] Under correct matching, the i-th target in camera 1 corresponds to two sets of candidate values, namely

[0093]

[0094] In these two sets of data, only one of each set is the result of correct matching, and under the condition of no error, the data set There is a value and data set If one of the values ​​in is exactly equal, the equal values ​​are considered a correct match. Then the i-th target in camera 1 is correctly matched with the j-th target in camera 2 and the k-th target in camera 3. In practice, due to the existence of errors, it is impossible to Therefore, we find The situation is considered a correct match. Then, update the matrix as follows

[0095]

[0096] The meaning of formula (12) is and The row and column are set to infinity so that they will not be included in the calculation when the next target is matched. In short, the matching target is The distance between the two values ​​is the smallest (due to the existence of errors, it may be the second smallest), and n sets of such values ​​are found to minimize the sum of the distances. Each set of such pairings is the optimal match. Note that the values ​​in each row and column of each matrix can only be used once.

[0097] After updating the matrix, repeat the above steps to get the correct matches for all other targets.

[0098] For the sake of intuitive understanding, the present invention assumes that n=3, that is, each matrix is ​​3×3. Specifically, it is as follows:

[0099] Step 2.1, first perform a match search for the first target in camera 1. The first target match is as follows: Fig.10 shown.

[0100] Step 2.2, after obtaining the best match between the first target in camera 1 and the targets in cameras 2 and 3, update the matrix as follows. Here we assume that the first target in camera 1 matches the first target in camera 2 and the third target in camera 3. After obtaining the first target match, update the matrix as follows Fig.11 shown.

[0101] Step 2.3: Repeat steps 2.1 and 2.2 until all matches are completed. The minimum sum of distances between all matches is the global optimum.

[0102] However, each time the matrix is ​​updated after each step, the matching of the next target may be non-optimal (because a row and a column are removed after each update). Can we find a globally optimal method to achieve multi-target matching?

[0103] Here, the present invention establishes a flexible step-by-step optimal method, that is, in each step, the first c matches with the smallest distance are selected (when the remaining targets are not less than c; if less than, the distance of the current number is the smallest). In this way, the step-by-step optimal process becomes a flexible step-by-step optimal process. Each optimization process is as follows Figure 4 shown.

[0104] Moreover, when selecting the c matches with the smallest distance, a distance threshold δ is set. If it is greater than this value, the selection is abandoned. Theoretically, the two best matches are selected each time, resulting in many final results, with a maximum of 2. n-1 results; however, since the matrix is ​​updated after the second minimum distance is selected, if the match is not the last match, the minimum distance after the subsequent matches will be greater than the threshold, especially after multiple steps of an incorrect match. Therefore, the final results will not be too many. The present invention adds the distance of each match on each result matching chain, and the one with the minimum distance is the optimal match. The distance addition in each update is as follows: Figure 5 shown.

[0105] The match with the minimum distance sum is the best match for multiple targets. If each distance corresponds to two values ​​in the two distance matrices Then the spatial position of the i-th target in camera 1 at height h can be expressed as

[0106]

[0107] The meaning of this formula is that the spatial position is the average of two spatial positions measured by camera 1 and camera 2,3 respectively.

[0108] Step 3: Position estimation at a specific moment.

[0109] The position of the target in camera 2 at each moment is measured by line-surface intersection. Now the matching is completed, so here only cameras 1 and 2 are used to estimate the position of a single target. After each single target position is estimated, the multi-target position estimation can be obtained. The geometric structure is as follows Figure 6 shown.

[0110] Figure 6 In the example, at time t, camera 2 captures the target and forms an image at x. t (u t v t ); Camera 1 does not capture the target at this moment, but captures the target at other moments. According to Formula 3, a plane S is obtained. According to the imaging x t, it can be determined that the target is located on the straight line l at time t t The expression of the straight line is as follows.

[0111]

[0112] Where a is an arbitrary scale factor, R c2 ,t c2 are the rotation matrix and translation vector of camera c2, respectively, obtained by PnP algorithm. Combining formula (3) and (14), we can get the position of the target at time t, which is the line l t The intersection point with plane S.

[0113] Example 1

[0114] Use three cameras to shoot the building, and the distance between the camera and the center of the public field of view is about 8 meters, as shown below Figure 7 shown.

[0115] The frame rate of the three cameras is 1000fps, the resolution is 1920×1080, the pixel size is 10μm, and the focal length is 20mm. First, the feature points are measured using a total station, and then the camera is calibrated using the feature points. The gaps between the wall tiles are simulated as linear motion trajectories, and the intersections between the gaps are the locations of the targets at each moment, such as Figure 8 shown.

[0116] The features at different heights on the trajectory are captured by only one camera, which means that the cameras are not synchronized when capturing moving objects, e.g. Figure 8 As shown. There are 8 trajectories in total, indicating that there are 8 moving objects. The intersection of the horizontal and vertical gaps between the bricks indicates the position of the object at a certain moment. The simulation scene is completed. Finally, the present invention is used to perform multi-target matching and spatial position estimation. After each matching and position estimation, the camera posture is adjusted and these steps are repeated in the next experiment. In one case, three cameras captured an image such as Fig. 9 shown.

[0117] For each test, we recorded the matching accuracy, and conducted 20 tests in total, and the matching rate of the method proposed in the present invention was 93.8%. After matching, the position was calculated, and the position error was 0.024m, indicating that the matching accuracy and position estimation accuracy of the present invention for moving multi-targets in asynchronous image sequences are both very high.

[0118] The above contents are further detailed descriptions of the present invention in combination with specific implementation methods, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.

Claims

1. A method for matching and estimating moving multiple targets based on asynchronous image sequences, characterized in that: The specific steps include: Step 1: Single target position estimation based on asynchronous image sequences Assuming the target moves in a straight line, the two cameras capture asynchronous images of the target; Assume that the left camera and the right camera capture multiple frames of target images respectively, and both the left camera and the right camera are calibrated. The conversion relationship between the camera coordinate system Oc1_Xc1Yc1Zc1, Oc2_Xc2Yc2Zc2 and the world coordinate system Ow_XwYwZw is known, satisfying the following relationship: The above formula is expressed as a 3D point (X w Y w Z w ) is projected on the left camera as (u c1 v c1 ); Among them, Z c1 is the scale factor, unknown; M c1 is the projection matrix of the left camera, which is obtained through prior calibration and is known; The left camera captures images of multiple targets, and a straight line is fitted on the image plane as follows: Among them (a c1 b c1 ) is the coefficient of the fitted straight line; Then the expression of the left camera plane where the target is located is as follows The expression of the right camera plane where the target is located based on the multi-frame target imaging position of the right camera is as follows: Among them, M c2 is the projection matrix of the right camera, obtained through prior calibration and known; (a c2 b c2 ) is the coefficient of the fitting line on the right camera image plane, and the target trajectory is the intersection of the left and right planes; the spatial position is used as the target matching information, and the target position is the intersection of a straight line and a plane. The calculation formula is as follows: f c1 (X w Y w Z w )=0 f c2 (X w Y w Z w )=0 (5) Z w =h In this way, the position estimation of a single target is achieved, and h is the height of the target; Step 2: Multi-target matching and position estimation at a specific height Assume that there are n targets entering the field of view, and use multiple frames of images to fit n straight lines on the image plane of camera 1. The expression of the nth straight line is as follows: Similarly, n straight lines are fitted on the image plane of camera 2 The expression of the nth straight line is as follows: Fitting the two sets of straight lines Treat it as a blind box and perform indiscriminate matching, that is, the i-th fitting line l in camera 1 i and all fitted lines in camera 2 Match them once respectively, 1≤i≤n, and then use the single target position estimation method in step 1 to obtain the candidate value of the i-th target position in camera 1 after each match: here represents the candidate value after the i-th fitting line in camera 1 and all the fitting lines in camera 2 are matched one by one, represents the target position after the i-th fitting line in camera 1 and the j-th fitting line in camera 2 are matched; perform the above operation on each fitting line in camera 1 to obtain a distance candidate matrix as follows: Introduce the third camera, and then similarly obtain a distance candidate matrix for camera 1 and camera 3, as follows: Under correct matching, the i-th target in camera 1 corresponds to two sets of candidate values, namely In these two sets of data, only one of each set is the result of correct matching, and under the condition of no error, the data set There is a value and data set in The values ​​in are exactly equal, and the equal values ​​are the correct match. The situation is considered a correct match; Update the matrix as follows Matching target is The distance between the two values ​​is the smallest, and n groups of such values ​​are found to minimize the sum of the distances. Each such pairing is the optimal match. After updating the matrix, repeat the above steps to obtain the correct matches for all other targets. Step 3: Position estimation at a specific moment The position of the target in camera 2 at each moment is measured by line-surface intersection. Now the matching is completed, so here only cameras 1 and 2 are used to estimate the position of a single target. At time t, camera 2 captures the target and forms an image at x. t (u t v t ); Camera 1 did not capture the target at this moment, but captured the target at other moments; According to Formula 3, a plane S is obtained, according to the imaging x t , it can be determined that the target is located on the straight line l at time t t The expression of the straight line is as follows: Where a is an arbitrary scale factor, R c2 ,t c2 are the rotation matrix and translation vector of camera c2 respectively, which are obtained by PnP algorithm. Combining formula (3) and (14), we can get the position of the target at time t, which is the straight line l t The intersection point with plane S.

2. The method for moving multi-target matching and position estimation based on asynchronous image sequences according to claim 1, characterized in that: In the formula (12) of step 2, the value of each row and column in each matrix can only be used once.

3. The method for moving multi-target matching and position estimation based on asynchronous image sequences according to claim 1, characterized in that: For the optimal match of step 2, a flexible step-by-step optimal method is established, that is, in each step, the first c matches with the smallest distance are selected. If the distance is lower than the current number, the step-by-step optimal process becomes a flexible step-by-step optimal process. When the c matches with the smallest distance are selected, a distance threshold δ is set. If it is greater than this value, the selection is abandoned. If each distance corresponds to two values ​​in the two distance matrices Then the spatial position P of the i-th target in camera 1 at height h is i Expressed as That is, the spatial position is the average of the two spatial positions measured by camera 1 and camera 2, 3 respectively; the distance of each match on each result matching chain is added, the one with the smallest distance is the optimal match, and the match with the minimum distance sum is the multi-target optimal match.

4. The method for matching and estimating moving multiple targets based on asynchronous image sequences according to claim 1, characterized in that: Assuming n=3, that is, each matrix is ​​3×3, the specific implementation process of step 2 is as follows: Step 2.1, first perform a match search for the first target in camera 1, and the first target match is shown in FIG10 ; Step 2.2, after obtaining the best match between the first target in camera 1 and cameras 2 and 3, the matrix is ​​updated as follows; assuming that the first target in camera 1 matches the first target in camera 2 and the third target in camera 3, the updated matrix after obtaining the first target match is shown in Figure 11; Step 2.3: Repeat steps 2.1 and 2.2 until all matches are completed; the minimum sum of distances between all matches is the global optimum.

Images