A target association recognition method based on multi-frame images

Abstract

This invention belongs to the field of target recognition technology and discloses a target association recognition method based on multi-frame images, including the following steps: S1, determining the search starting point; S2, based on the determined starting trajectory, selecting two target points on one of the trajectories to predict the target position in the third frame; S3, starting the next process, searching from the second frame, with the coordinates of all targets in the second frame as the starting point, referring to S1 to S2, until the coordinates of all targets in the Nth frame are the starting point; S4, deleting duplicate false alarm tracks. This invention performs data pruning during the depth-first search process, pruning data not in the velocity gate and excluding it from the least squares estimation calculation, while also pruning erroneous tracks. This not only avoids the curse of dimensionality but also greatly improves computational power under high false alarm conditions.

Description

Technical Field

[0001] This invention belongs to the field of target recognition technology, specifically relating to a target association recognition method based on multi-frame images. Background Technology

[0002] The multi-frame images described in this invention refer to a continuous set of images acquired by an image / video acquisition device over a period of time, and the content or main target in the images is in an overall translational motion state, including dynamic video frames and time-series monitoring images of optical devices.

[0003] Target tracking, as an important aspect of computer science and technology development, has a long history, dating back to the eve of World War II, when tracking systems were mostly one-to-one. With the continuous development of target detection technology, the possibility of detecting multiple targets simultaneously has increased, bringing many new challenges to detection and tracking.

[0004] The purpose of multi-target tracking is to decompose the received measurement data into different sets of observations or trajectories corresponding to different information sources. In multi-frame image target association and recognition, it is necessary to traverse all points of the target in each frame. When the amount of data is large, traversal and search will lead to the curse of dimensionality. The main reason for the curse of dimensionality is that there are too many invalid false alarm points in the data. Summary of the Invention

[0005] The purpose of this invention is to propose a target association recognition method based on multi-frame images. By utilizing the changes of targets in the time domain and the correlation between adjacent frames in multi-frame images, and based on the correspondence between the previous frame and the current frame, the motion information of objects between adjacent frames is calculated, thereby realizing the association recognition of image sequences, extracting the motion trajectory of moving targets, and eliminating false target points.

[0006] The technical solution adopted in this invention is a target association recognition method based on multi-frame images, comprising the following steps:

[0007] S1, Determine the starting point of the search;

[0008] S2, based on the determined starting trajectory, select two target points on one of the trajectories to predict the target position in the third frame;

[0009] S3, begin the next process, start the search from the second frame, with the coordinates of all targets in the second frame as the starting point, refer to S1 to S2, until the coordinates of all targets in the Nth frame are the starting point;

[0010] S4, delete duplicate false alarm tracks.

[0011] Furthermore, in S1, determining the search starting point specifically involves:

[0012] The search begins from the first frame, with the coordinates of all targets in the first frame serving as the starting point. A speed window L1 is set based on the exposure time, and P1 represents the targets in the first frame. The search then continues within the speed window of P1 to find targets that fall within the window in the second frame.

[0013] If no target falls in the second frame, prune the trail, delete the trajectory, and start searching for other target points from the first frame.

[0014] If P2 is a target in the second frame and falls within the velocity window L1 of P1, record all the required combinations of all target points in the first frame and the starting points falling within the velocity window L1 as the starting trajectory.

[0015] Furthermore, in S2, based on the determined initial trajectory, selecting two target points from one of the trajectories to predict the target position in the third frame specifically involves:

[0016] Based on the starting trajectory and timing determined from the target points in the first and second frames, two target points from one of the trajectories are selected to predict the position of target P3 in the third frame. The radius of the detection window L2 in the nth frame is also considered. Where n≥3, ε is the systematic error, and there are three cases during detection:

[0017] If no target point falls within the detection window L2, the trajectory is pruned and deleted, the depth-first search starting from that point is terminated, and the depth-first search starts from other trajectories. This process continues until the Nth frame, where N is the total number of frames.

[0018] If only one target point falls within the detection window L2, then this point is included in the trajectory. Then, the updated trajectory data is used to predict the target point position in the next frame using the least squares method and to calculate the detection window size in the next frame.

[0019] If multiple target points fall within the detection window L2, the center point of the spatial target is estimated using the first moment and included in the trajectory. Then, the target point position in the next frame is predicted using the least squares method and the detection window size in the next frame is calculated using the updated trajectory data.

[0020] Furthermore, in step S4, deleting duplicate false alarm tracks specifically involves:

[0021] Among multiple tracks starting from the same point, the trajectory curve is fitted using the track position, and the error of each track is calculated using the associated target point position and trajectory curve. The track with the smallest error is retained, and the other tracks are deleted.

[0022] The beneficial effects of this invention are as follows:

[0023] Data pruning is performed during the depth-first search process, pruning data that is not in the velocity gate and therefore not participating in the least squares estimation calculation. At the same time, erroneous tracks are also pruned, which not only avoids the curse of dimensionality but also greatly improves the computing power under high false alarm conditions. Attached Figure Description

[0024] Figure 1 This is a flowchart of the present invention;

[0025] Figure 2 This is a schematic diagram illustrating the process of determining the search starting point in this invention;

[0026] Figure 3 A schematic diagram of the flight paths triggered by false alarm points;

[0027] Figure 4 This is a two-dimensional display diagram of the association recognition results of the present invention;

[0028] Figure 5 This is a three-dimensional representation of the association recognition results of the present invention;

[0029] Wherein, L1 is the velocity window; L2 is the detection window; p1 is the target in the first frame; p2 is the target in the second frame; p3 is the target in the third frame; S is the trajectory; A is the real target; and B is the false alarm point. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this invention clearer and more understandable, the technical solutions of this invention will be clearly and completely described below in conjunction with the accompanying drawings.

[0031] This invention proposes to use a straight-line detection method for target association recognition in multi-frame images. The motion of spatial targets within a short time frame can be considered uniform linear motion, requiring estimation of motion parameters. For uniform linear motion, the least squares method is the optimal parameter estimation method. However, with large datasets, traversal searching can lead to the curse of dimensionality, primarily due to an excessive number of invalid false alarms. An effective solution is to perform data pruning during the depth-first search process. Data not in the velocity gate is pruned and excluded from the least squares estimation calculation. Incorrect tracks are also pruned. This not only avoids the curse of dimensionality but also significantly improves computational power under high false alarm conditions. Figure 1 , Figure 2 As shown, the specific steps include:

[0032] S1, Determine the starting point of the search;

[0033] The process of determining the search starting point is also the process of determining the starting trajectory. Specifically, the search begins from the first frame. The coordinates of all targets (noise, lens blemishes, false alarm targets, real targets, etc.) in the first frame are used as the starting point. A speed window L1 is set according to the exposure time, and P1 represents the target in the first frame. The search is then conducted within the speed window P1 to find targets that fall within the window in the second frame.

[0034] If no target falls in the second frame, prune the trajectory, delete it, and start searching for other target points from the first frame.

[0035] If P2 is a target in the second frame (and also a target in the first frame) and falls within the velocity window L1 of P1, record all the required combinations of all target points in the first frame and the starting points falling within the velocity window L1 as the starting trajectory.

[0036] S2, based on the determined starting trajectory, select two target points on one of the trajectories to predict the target position in the third frame;

[0037] Specifically, based on the starting trajectory determined by the target points in the first and second frames and the time sequence, one trajectory is selected, such as the two target points of trajectory S, to predict the position of target P3 in the third frame (P3 is also a target in both the second and third frames). The radius of the detection window L2 in the nth (n≥3) frame is... Where ε is the systematic error, there are three cases during detection:

[0038] If no target point falls within the detection window L2, the trajectory is pruned and deleted, the depth-first search starting from that point is terminated, and a depth-first search is started from another trajectory. This process continues until the Nth frame, where N is the total number of frames.

[0039] If only one target point falls within the detection window L2, this point is included in the trajectory. Then, the updated trajectory data is used to predict the target point position in the next frame using the least squares method, and the size of the detection window in the next frame is calculated.

[0040] If multiple target points fall within the detection window L2, this situation is unlikely to occur in most cases because the target is usually much larger than the detection window L2, making it difficult for this situation to occur. If it does occur, the center point of the spatial target is estimated using the first moment, and this point is included in the trajectory. Then, the updated trajectory data is used to predict the target point position in the next frame using the least squares method, and the size of the detection window in the next frame is calculated.

[0041] S3, begin the next process, starting the search from the second frame. The coordinates of all targets in the second frame are the starting points. Refer to S1 to S2, until the coordinates of all targets in the Nth frame are the starting points.

[0042] S4, delete duplicate false alarm tracks;

[0043] After obtaining the track through correlation identification, some false alarm points may be too close to the track, thus inducing the generation of false alarm tracks, such as... Figure 3 As shown, this can lead to two or more false alarm tracks starting from a single point. Most of these false alarm tracks share the same trajectory, differing only in a few starting points. Therefore, after obtaining the tracks, tracks with the same starting point are merged and deleted. Specifically, among multiple tracks starting from the same point, a trajectory curve is fitted using the track position. Then, the error of each track is calculated using the associated target point position and the trajectory curve. The track with the smallest error is retained, and the others are deleted.

[0044] Figure 4 It is the superposition of the target point's trajectory on different frames, which can show the target point's motion trajectory; Figure 5 It displays the relationship between the motion trajectory of the target point in different frames, as well as the entry and removal of the target point.

[0045] All content not described in detail in this specification is well-known to those skilled in the art.

Claims

1. A target association and recognition method based on multi-frame images, characterized in that, Includes the following steps: S1, Determine the starting point of the search; S2, based on the determined initial trajectory, select two target points from one of the trajectories to predict the target position in the third frame, specifically: Based on the starting trajectory and timing determined by the target points in the first and second frames, two target points from one of the trajectories are selected to predict the position of the target (P3) in the third frame. The radius of the detection window (L2) in the nth frame is... Where n≥3, ε is the systematic error, and there are three cases during detection: If no target point falls within the detection window (L2), the trajectory is pruned and deleted, the depth-first search starting from that point is terminated, and the depth-first search starts from other trajectories. This process continues until the Nth frame, where N is the total number of frames. If only one target point falls within the detection window (L2), then this point is included in the trajectory. Then, the updated trajectory data is used to predict the target point position in the next frame using the least squares method and to calculate the detection window size in the next frame. If multiple target points fall within the detection window (L2), the center point of the spatial target is estimated using the first moment, and this point is included in the trajectory. Then, the target point position in the next frame is predicted using the least squares method and the detection window size in the next frame is calculated using the updated trajectory data. S3, begin the next process, start the search from the second frame, with the coordinates of all targets in the second frame as the starting point, refer to S1 to S2, until the coordinates of all targets in the Nth frame are the starting point; S4, delete duplicate false alarm tracks.

2. The target association recognition method based on multi-frame images according to claim 1, characterized in that, In S1, determining the search starting point specifically involves: The search begins from the first frame, with the coordinates of all targets in the first frame serving as the starting point. A speed window (L1) is set based on the exposure time, and (P1) represents the targets in the first frame. Within the speed window (P1), targets falling within the window in the second frame are searched. If no target falls in the second frame, prune the trail, delete the trajectory, and start searching for other target points from the first frame. If there exists (P2) as a target in the second frame and it falls within the velocity window (L1) of (P1), record all the required combinations of all target points in the first frame and the starting points falling within the velocity window (L1) as the starting trajectory.

3. The target association recognition method based on multi-frame images according to claim 1, characterized in that, In step S4, deleting duplicate false alarm tracks specifically involves: Among multiple tracks starting from the same point, the trajectory curve is fitted using the track position, and the error of each track is calculated using the associated target point position and trajectory curve. The track with the smallest error is retained, and the other tracks are deleted.

Images