A moving target recognition and tracking method and readable storage medium

Abstract

The application provides a moving target recognition and tracking method, which comprises the following steps: uniformly distributing particles on initial video images of current continuous N frames according to particle simulation motion states to obtain images to be recognized of the current continuous N frames; obtaining a best matching area of a target image in the images to be recognized of the current continuous N frames; correcting the particles in the target image according to the best matching area of the target image to obtain particle weights of the target image; resampling the target image according to the particle weights corresponding to each particle; and obtaining a moving target position in the target image according to the particle distribution after resampling of the target image. The application solves the problem of the moving target recognition and tracking method in the prior art that has low robustness and poor dynamic recognition effect, and improves the robustness and dynamic recognition effect of the moving target recognition and tracking.

Description

Technical Field

[0001] This invention relates to the field of moving target tracking technology, and in particular to a moving target identification and tracking method and a readable storage medium. Background Technology

[0002] Moving target recognition and tracking is an important research direction in the fields of preventing theft and damage to immovable cultural relics and video security monitoring. It is a fundamental and key technology for applications such as visual navigation of mobile robots, intelligent monitoring, precision weapon guidance, and human-computer interaction systems.

[0003] Currently, moving target recognition and tracking commonly employs methods such as infrared camera sensing, fixed background comparison, optical flow field analysis algorithms, and traditional particle filtering algorithms. Infrared imaging technology utilizes infrared cameras of different wavelengths to image scenes containing targets, and then extracts and senses targets within the scene based on the physical characteristics of the infrared images. Using a parametric model of the background image to approximate pixel values, the current frame is compared with the background image using difference to detect moving regions. Pixel regions with significant differences are considered moving regions, while those with smaller differences are considered background regions. Optical flow primarily represents motion in image brightness patterns, while optical flow field refers to surface motion in grayscale patterns. Under normal conditions, we can estimate relative motion based on image motion. Optical flow field analysis algorithms establish displacement vectors between adjacent frames and calculate the optical flow of the global image. In traditional particle filtering algorithms, a set of samples (or particles) is used to approximate the posterior probability distribution of the system, and this approximate representation is then used to estimate the state of the nonlinear system. Furthermore, to reduce particle degradation, resampling is performed, i.e., more particles are scattered near particles with high weights, while particles with low weights are removed.

[0004] Infrared cameras can only detect targets with temperature characteristics, such as people, vehicles, and animals, and have limitations in identifying moving objects and drones. Fixed-background comparison methods are ineffective for identifying targets in rapidly changing backgrounds. Optical flow field analysis algorithms are computationally expensive and lack robustness due to interference from light, wind, and occlusion. Traditional particle filter tracking algorithms resample based on particle weights, directly removing low-weight particles. This can lead to the discarding of correct but low-weight particles, resulting in particle diversity degradation. This makes them ineffective and lacks robustness when dealing with drastic changes in target posture or sudden disappearances.

[0005] Existing moving target recognition and tracking methods suffer from poor robustness and unsatisfactory dynamic recognition results. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method for recognizing and tracking moving targets, which solves the problems of weak robustness and unsatisfactory dynamic recognition results in existing moving target recognition and tracking methods.

[0007] In a first aspect, embodiments of the present invention provide a method for recognizing and tracking moving targets, comprising: uniformly distributing particles on current consecutive N frames of initial video images according to the particle's simulated motion state to obtain current consecutive N frames of images to be recognized; obtaining the best matching region of a target image in the current consecutive N frames of images to be recognized; wherein, N>0, and the target image is one of the middle frames in the current consecutive N frames of images to be recognized; correcting the particles in the target image according to the best matching region of the target image to obtain the particle weights of the target image; resampling the target image according to the particle weights corresponding to each particle; and obtaining the position of the moving target in the target image according to the particle distribution after resampling of the target image.

[0008] Optionally, when N=3, obtaining the best matching region of the target image in the current consecutive N frames of images to be identified includes: outlining the target on the first frame image and establishing a target template for the first frame image; determining the starting search center of the third frame image based on vertical and horizontal parameters, and starting the search from the starting search center using the search template to obtain the best matching region of the third frame image; deriving the first starting search center of the second frame image using a particle filter algorithm based on the first frame image and the target template; deriving the second starting search center of the second frame image using a particle filter algorithm based on the third frame image and the best matching region of the third frame image; comparing the first starting search center and the second starting search center, selecting a new center point region where they overlap, and starting the search from the new center point region using the search template to obtain the best matching region of the second frame image; wherein, the second frame image is used as the target image.

[0009] Optionally, based on the first frame image and the target template, the first starting search center of the second frame image is derived using a particle filtering algorithm, including: determining the state transition equation based on the first frame image and the target template, substituting each particle in the first frame image into the state transition equation to obtain the predicted particle corresponding to the second frame image, where the state transition equation is: x k =f(x) k-1 )+u k-1 , where x k Let x be the particle state at the current moment. k-1 Let f be the particle state at the previous moment, and f be the state transition function. k-1 The process noise is used to obtain the first starting search center of the second frame image based on the predicted particles.

[0010] Optionally, based on the third frame image and its optimal matching region, a particle filtering algorithm is used to deduce the second starting search center of the second frame image. This includes: determining a state transition equation based on the third frame image and its optimal matching region; substituting each particle in the third frame image into the state transition equation; and deduce the particle state of the second frame image. The state transition equation is: x k =f(x) k-1 )+u k-1 , where x k-1 Let x be the particle state at the current moment. k Let f be the particle state at the next moment, and let u be the state transition function. k-1 This is process noise.

[0011] Optionally, the search template is a circular search template.

[0012] Optionally, the search for the optimal matching region of the third frame image is initiated at the starting search center using a search template. This includes: firstly searching in the vertical and horizontal directions at the starting search center using the search template, selecting a first minimum matching value as the optimal matching value, and accumulating the occurrence of a first location region; determining the optimal matching region based on the first location region; continuing the search in other directions at the starting search center using the search template, and if a matching value smaller than the first minimum matching value is found, updating the optimal matching value to a smaller matching value, accumulating the occurrence of a new location region near the smaller matching value, and using the new location region as the optimal matching region; if a matching value smaller than the optimal matching value is found, continuing to update the optimal matching value and the optimal matching region until the search is completed.

[0013] Optionally, based on the best matching region of the target image, the particles in the target image are corrected to obtain the particle weights of the target image, including: obtaining the observed values ​​of the particles based on the best matching region of the target image, wherein the observation equation is: y k =h(x k )+v k Where h is the observation equation, x k Let y be the particle state at time k. k For the observed value, v k To reduce observation noise; obtain the observation value y at time k. k Then, the prior probability density is updated using Bayes' theorem to obtain the posterior probability density, and the posterior probability density equation is: in Y is the normalization constant. k Y represents all observations from time 0 to time k. k-1p(x) represents all observations from time 0 to k-1. k |Y k-1 Let p(x) be the prior probability density. k |Y K ) represents the posterior probability density; the target image particles are corrected based on the posterior probability density, and the target image particle weights are updated.

[0014] Optionally, the target image is resampled according to the particle weight corresponding to each particle, including: classifying particles with large weights as main particles and particles with small weights as secondary particles; scattering particles in the vicinity of the main particles in descending order of their weights; retaining some secondary particles, and scattering fewer or no particles in the vicinity of the retained secondary particles.

[0015] Optionally, obtaining the position of the moving target in the target image based on the particle distribution after resampling the target image includes: performing curve fitting on the trajectory of the moving target based on the least squares method according to the principal particle distribution after resampling the target image; and obtaining the position of the moving target in the target image based on the fitted trajectory of the moving target.

[0016] Secondly, embodiments of the present invention provide a computer-readable storage medium storing at least one instruction, which is executed by a processor to implement the steps of the above-described method.

[0017] The technical principle of this invention is as follows: By continuously acquiring the current N frames of initial video images, a large number of particles are evenly scattered in each frame to simulate motion, resulting in a current series of N consecutive frames of images to be identified, uniformly filled with particles. The target image is one frame from the current series of N consecutive frames of images to be identified, excluding the first and last frames. Based on the current series of N consecutive frames of images to be identified, the optimal matching region of the target image is obtained by combining forward deduction from images before the target image and backward deduction from images after the target image. The forward deduction result and the backward deduction result are compared to obtain an optimal region as the optimal matching region of the target image. Then, based on the optimal matching region of the target image, the particle weights of the target image are corrected to obtain new particle weights. The target image is then resampled according to the particle weights corresponding to each particle, i.e., the particle distribution is rearranged. Based on the particle distribution after resampling of the target image, the position of the moving target in the target image is obtained.

[0018] Compared with the prior art, the present invention has the following beneficial effects:

[0019] By employing the method of fusing and iteratively determining the optimal matching region of the target image using N consecutive initial video images, the one-way derivation mode between consecutive frames is broken. Starting from the first and last frames, a cross-coupled iterative mode is adopted to quickly derive the central particle that has the greatest impact on the moving target, and rapidly derive the tracking result of the target image. This effectively overcomes the influence of light interference, environmental wind, occlusion, etc., and achieves reliable identification and tracking of moving targets, improving the real-time performance and robustness of moving target identification and tracking. Attached Figure Description

[0020] Figure 1 A schematic diagram illustrating a moving target recognition and tracking method provided in an embodiment of the present invention;

[0021] Figure 2 This is a schematic diagram illustrating the acquisition of the best matching region of the target image in three consecutive frames of images to be identified, provided by an embodiment of the present invention.

[0022] Figure 3 A schematic diagram illustrating the steps of performing a search using the search template provided in this embodiment of the invention;

[0023] Figure 4 This is a schematic diagram illustrating the resampling of a target image according to an embodiment of the present invention. Detailed Implementation

[0024] The technical solutions of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0025] like Figure 1 As shown, this embodiment of the invention provides a method for recognizing and tracking moving targets, which includes the following steps:

[0026] S100: Based on the simulated motion state of the particles, the particles are evenly distributed on the current consecutive N frames of initial video images to obtain the current consecutive N frames of images to be identified.

[0027] S200, obtain the best matching region of the target image in the current consecutive N frames of images to be identified; wherein the target image is one of the frames in the current consecutive N frames of images to be identified except for the first frame and the last frame, N≥3 and is a positive integer.

[0028] S300, based on the best matching region of the target image, the particles in the target image are corrected to obtain the particle weights of the target image.

[0029] S400, the target image is resampled according to the particle weight corresponding to each particle.

[0030] S500: Based on the particle distribution after resampling the target image, obtain the position of the moving target in the target image.

[0031] The detailed working process of this embodiment is as follows: N initial video images are continuously acquired. A large number of particles are evenly distributed in each frame to simulate motion, resulting in a continuous N-frame image to be identified, uniformly filled with particles. The target image is one of the N frames excluding the first and last frames. Based on the N frames, the optimal matching region for the target image is obtained by combining forward deduction from images preceding the target image and backward deduction from images following the target image. The forward and backward deduction results are compared to obtain an optimal region as the optimal matching region for the target image. Then, based on the optimal matching region, the particle weights of the target image are corrected to obtain new particle weights. The target image is resampled according to the particle weights corresponding to each particle, i.e., the particle distribution is rearranged. Based on the resampled particle distribution of the target image, the position of the moving target in the target image is obtained. By employing an iterative method that fuses N consecutive initial video frames to determine the optimal matching region of the target image, this approach breaks the unidirectional derivation mode between consecutive frames. Starting with the first and last frames, a cross-coupling iterative mode is used to rapidly derive the central particle that has the greatest impact on the moving target, quickly deriving the tracking results of the target image. This effectively overcomes the effects of light interference, environmental wind, and occlusion, achieving reliable identification and tracking of moving targets. It improves the real-time performance and robustness of moving target identification and tracking, as well as the dynamic recognition effect. It enables rapid measurement and automatic identification of multiple moving targets, real-time processing of digital video, and automatic completion of moving target detection, identification, and tracking. The recognition frame rate meets the requirements for real-time alarms. The research results can be widely used in the fields of anti-theft and anti-vandalism video security for immovable cultural relics, and can also be extended to areas such as key site security, drone protection, and precision guidance and navigation.

[0032] like Figure 2 As shown in the embodiment of the present invention, when N=3, obtaining the best matching region of the target image in the current consecutive N frames of images to be identified includes the following steps:

[0033] S210, outline the target on the first frame image and create a target template for the first frame image.

[0034] S220: Based on the vertical and horizontal parameters, determine the starting search center of the third frame image, and start searching at the starting search center using the search template to obtain the best matching area of ​​the third frame image.

[0035] S230, based on the first frame image and the target template, derive the first starting search center of the second frame image using a particle filter algorithm.

[0036] S240, based on the third frame image and the best matching region of the third frame image, the particle filter algorithm is used to deduce the second starting search center of the second frame image.

[0037] S250, compare the first starting search center and the second starting search center, select a new center point region where they overlap, and start searching from the new center point region using a search template to obtain the best matching region of the second frame image; wherein, the second frame image is used as the target image.

[0038] The detailed working process of this embodiment is as follows: Let N equal 3, and using the second frame image as the target image, the size of the first frame image is L*L. The target's position is outlined on the first frame image, and this outlined target is used as the target template for subsequent template matching. Based on the calculated vertical and horizontal parameters (which are the coordinate parameters around the target), the starting search center of the search template in the third frame image is determined. The search begins from this starting search center using the search template, which is a box with a certain shape and size. The image region corresponding to the search template is matched with the target template, and the region in the image most similar to the target template is found as the best matching region in the third frame image. Based on the first frame image and the target template within it, a particle filter algorithm is used to derive the first initial search center for the second frame image. This involves using the initially uniformly distributed particles in the first frame image as random variables to construct a probabilistic model predicting the probability of the target template's particle swarm appearing at the next moment. This model estimates the solution to the problem and predicts the target template's location region in the second frame image. A region is then selected as the first initial search center for the second frame image. Based on the third frame image and its optimal matching region, a particle filter algorithm is used to reverse-engineer the second initial search center for the second frame image. This involves using the particle swarm near the optimal matching region in the third frame image as the predicted location of the target's particle swarm from the previous moment. The probabilistic model is used to reverse-engineer the target's particle swarm location region from the previous moment (i.e., the second frame image), and a region is selected as the second initial search center. The first and second initial search centers are compared, and new center point regions are set in 3-5 overlapping directions. The search template is then used to search from these new center point regions to obtain the optimal matching region for the second frame image.

[0039] By employing a search template to begin the search from the initial search center, the computational load is reduced. This differs from traditional particle filter target tracking algorithms, which use an exhaustive search method, scanning all pixels of candidate modules based on the target module, requiring recalculation with each movement, resulting in high computational costs. By combining particle filtering with template matching, the particle filtering algorithm first predicts the search center of the target image. Then, by comparing the search centers of two target images (forward and backward) from previous and subsequent frames, the optimal region is selected before further searching and template matching. This allows for rapid derivation of the central particle with the greatest impact on the moving target, quickly deriving the tracking result of the target image. Even when the target image is affected by light interference, environmental wind, occlusion, etc., the fusion results of forward and backward derivation from previous and subsequent frames can be used to correct the target's position, achieving reliable moving target identification and tracking. This improves the real-time performance, robustness, and dynamic recognition effectiveness of moving target identification and tracking.

[0040] In some embodiments, deriving the first initial search center of the second frame image using a particle filtering algorithm based on the first frame image and the target template includes: determining a state transition equation based on the first frame image and the target template; substituting each particle in the first frame image into the state transition equation to obtain the predicted particle corresponding to the second frame image; the state transition equation is: x k =f(x) k-1 )+u k-1 , where x k Let x be the particle state at the current moment. k-1 Let f be the particle state at the previous moment, and f be the state transition function. k-1 The process noise is used to predict the particle state. The first initial search center of the second frame image is obtained based on the predicted particles. The current state depends only on the previous state. The state transition equation is derived from the previous state, and it is based on a large amount of experimental data. In this way, the particle state at the current moment, i.e., the particle's location, can be predicted using particle filtering to determine the first initial search center of the second frame image, facilitating a more accurate and faster target search.

[0041] In some embodiments, based on the third frame image and the best matching region of the third frame image, a particle filtering algorithm is used to deduce the second starting search center of the second frame image, including: determining a state transition equation based on the third frame image and the best matching region of the third frame image; substituting each particle in the third frame image into the state transition equation; and deduce the particle state of the second frame image. The state transition equation is: x k =f(x) k-1 )+u k-1 , where xk-1 Let x be the particle state at the current moment. k Let f be the particle state at the next moment, and let u be the state transition function. k-1 This is process noise. By using the known state of the particle at the next moment and then using the state transition equation to deduce the particle state at the current moment, we can obtain the second starting search center of the second frame image in another way. This facilitates comparison with the first starting search center of the second frame image to find the optimal starting search center of the second frame image.

[0042] In some embodiments, the search template is a circular search template. In a moving context, a multimodal circular matching search algorithm is proposed to improve upon the traditional particle filter target tracking algorithm. A circular neighborhood replaces the traditional square neighborhood of pixel search, and the original 3*3 neighborhood is expanded to an arbitrary size neighborhood. The coverage area calculated in one operation is approximately three times that of the traditional square area, increasing the speed by about three times, thus meeting the requirements for fast search. Since the probability of a target appearing in all directions at the search point is equal, using the circular search template to determine the location of the optimal point can quickly approximate the precise region of the optimal matching point. The optimal matching point detected at this point is the iterative search result, enabling fast search and tracking of multiple moving targets.

[0043] like Figure 3 As shown, in an embodiment of the present invention, the best matching region of the third frame image is obtained by starting a search at the initial search center using a search template, including the following steps:

[0044] S221, using a search template, first search in the vertical and horizontal directions at the starting search center, select the first minimum matching value as the best matching value, and accumulate the first location area where it occurs.

[0045] S222, determine the optimal matching region based on the first location region.

[0046] S223, continue searching in other directions from the starting search center using the search template. If a matching value smaller than the first minimum matching value is found, update the best matching value to a smaller matching value and accumulate new location areas near the smaller matching value. Use the new location areas as the best matching areas. If a matching value smaller than the best matching value is found, continue updating the best matching value and the best matching areas until the search is complete.

[0047] The detailed working process of this embodiment is as follows: First, a search is performed in the vertical and horizontal directions using a search template at the initial search center. The smallest matching value is taken as the optimal matching value. The location area near the optimal matching value is then considered the optimal matching region. The matching value represents the degree of matching between the image region corresponding to the search template and the target template; the smaller the matching value, the closer it is to the target template. Then, the search module searches in other directions. When a smaller matching value is found, the optimal matching value and optimal matching region are iterated until the search is complete. The smallest matching value found in each direction is taken as the optimal matching value, and the location area near the smallest matching value found in each direction is taken as the optimal matching region for the third frame image. Because the optimal matching value and optimal matching region are iterated, the smallest matching value closest to the target template can be found in each direction, making the final result more accurate.

[0048] In some embodiments, the particles in the target image are corrected based on the best matching region of the target image to obtain the particle weights of the target image, including: obtaining the observed values ​​of the particles based on the best matching region of the target image, wherein the observation equation is: y k =h(x k )+v k Where h is the observation equation, x k Let y be the particle state at time k. k For the observed value, v k To reduce observation noise; obtain the observation value y at time k. k Then, the prior probability density is updated using Bayes' theorem to obtain the posterior probability density, and the posterior probability density equation is: in Y is the normalization constant. k Y represents all observations from time 0 to time k. k-1 p(x) represents all observations from time 0 to k-1. k |Y k-1 Let p(x) be the prior probability density. k |Y K The posterior probability density is used to correct the target image particles and update their weights. Particle filtering is a nearsighted Bayesian filtering algorithm based on Monte Carlo simulation. Its core idea is to approximate the probability density function of the system's random variable with a large number of sample points in the state space. Based on the best matching region closest to the target found by the search template, an observation equation can be established to obtain the observed values. The particle weights are then recalculated using the posterior probability density equation; the closer the value is to the observed value, the greater the weight. The previous weights were predicted, and this process corrects and normalizes the particle weights to facilitate subsequent resampling of the target image.

[0049] like Figure 4 As shown, in an embodiment of the present invention, resampling the target image according to the particle weight corresponding to each particle includes the following steps:

[0050] S410 classifies particles with higher weights as master particles and particles with lower weights as slave particles.

[0051] S420, in the vicinity of the main particle, releases particles in descending order of their weight according to the main particle's weight.

[0052] S430, retains a portion of the particles, and sprinkles fewer or no particles near the retained particles.

[0053] The detailed working process of this embodiment is as follows: First, particles with high weights are classified as master particles, and particles with low weights are classified as slave particles. Then, particles are re-scattered within the master particles according to their weights. More particles are scattered around particles with high weights (high similarity), and fewer particles are scattered around particles with low weights (low similarity). Some slave particles are retained, while others are removed. Particles can be scattered less or not at all near slave particles. This re-scattering of particles according to weights achieves particle resampling, highlighting the target positions corresponding to the new weights. By retaining some slave particles, particle diversity can be maintained. Before resampling, the weight of each particle is calculated. Sometimes, due to high environmental similarity or the influence of measurement noise, particles close to the correct state may have lower weights, while particles in the wrong state may have higher weights. Resampling is based on particle weights, so correct particles may be discarded. Frequent resampling further increases the possibility of discarding correct but low-weight particles. Particle diversity refers to the differences between particles. Initially, there are ten particles. If resampling occurs, five particles are discarded, and the remaining five are used to create five duplicate particles. At this point, only five of the ten particles are different, meaning particle diversity is reduced. While resampling can improve particle scarcity, it also decreases particle diversity. This embodiment retains some particles, specifically those with correct but low weights. During the next resampling, when recalculating weights, these particles can have their weights recalculated, returning them to a higher weight state. This reduces computational load while preserving particle diversity, leading to more accurate recognition results. Extensive training experiments were conducted using this method on various states, including single-target, multi-target, static background, and dynamic background. This improvement, to a certain extent, ensures particle diversity and demonstrates good performance in handling situations involving rapid target changes, partial occlusion, or temporary disappearance.

[0054] In some embodiments, obtaining the position of a moving target in the target image based on the particle distribution after resampling the target image includes: performing curve fitting on the trajectory of the moving target using the least squares method based on the principal particle distribution after resampling the target image; and obtaining the position of the moving target in the target image based on the fitted trajectory. Resampling distributes particles according to updated particle weights, resulting in a different particle distribution. Based on the resampling principal particle distribution, the target particle has the highest weight and the densest particle distribution. Curve fitting using the least squares method allows the position of the moving target to be observed from the fitted curve. Thus, curve fitting on the trajectory of the moving target using the least squares method greatly improves the control accuracy of the moving target trajectory position and further enhances the display effect of multiple moving target trajectories.

[0055] An embodiment of the present invention provides a computer-readable storage medium storing at least one instruction, which is executed by a processor to implement the steps of the above method.

[0056] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for recognizing and tracking moving targets, characterized in that, include: Based on the simulated motion state of the particles, the particles are evenly distributed on the current consecutive N frames of initial video images to obtain the current consecutive N frames of images to be identified. Obtain the best matching region of the target image in the current consecutive N frames of images to be identified; wherein the target image is one of the frames in the current consecutive N frames of images to be identified, excluding the first frame and the last frame; Based on the best matching region of the target image, the particles in the target image are corrected to obtain the particle weights of the target image; The target image is resampled according to the particle weight corresponding to each particle; Based on the particle distribution after resampling the target image, the position of the moving target in the target image is obtained; When N=3, the best matching region of the target image in the current N consecutive frames of images to be identified is obtained, including: The target is outlined in the first frame image, and a target template for the first frame image is created. Based on the vertical and horizontal parameters, the starting search center of the third frame image is determined, and the best matching area of ​​the third frame image is obtained by starting the search at the starting search center using the search template. Based on the first frame image and the target template, the first starting search center of the second frame image is derived using the particle filter algorithm; Based on the third frame image and the best matching region of the third frame image, the particle filter algorithm is used to deduce the second starting search center of the second frame image; The first starting search center and the second starting search center are compared, and a new center point region is selected where they overlap. The search template is used to start searching from the new center point region to obtain the best matching region of the second frame image. The second frame image is used as the target image.

2. The moving target recognition and tracking method as described in claim 1, characterized in that, Based on the first frame image and the target template, the first initial search center of the second frame image is derived using a particle filter algorithm, including: Based on the first frame image and the target template, the state transition equation is determined. Each particle in the first frame image is then substituted into the state transition equation to obtain the predicted particle for the second frame image. The state transition equation is as follows: ,in, The current state of the particle. The particle state at the previous moment. This is the state transition function. This is process noise; The first starting search center of the second frame image is obtained based on the predicted particles.

3. The moving target recognition and tracking method as described in claim 1, characterized in that, Based on the third frame image and its best matching region, the particle filter algorithm is used to deduce the second starting search center of the second frame image, including: Based on the third frame image and its optimal matching region, a state transition equation is determined. Each particle in the third frame image is then substituted into this equation to deduce the particle state in the second frame image. The state transition equation is: ,in, The current state of the particle. The particle state at the next moment. This is the state transition function. This is process noise.

4. The moving target recognition and tracking method as described in claim 1, characterized in that, The search template is a circular search template.

5. The moving target recognition and tracking method as described in claim 1, characterized in that, The search begins at the starting search center using a search template, yielding the best matching region for the third frame image, including: The search template is used to first search in the vertical and horizontal directions of the starting search center, select the first minimum matching value as the best matching value, and accumulate the first location area where the match occurs. The optimal matching region is determined based on the first location region; The search continues in other directions from the starting search center using the search template. If a match value smaller than the first minimum match value is found, the best match value is updated to a smaller match value, and a new location region is accumulated near the smaller match value. This new location region is taken as the best match region. If a match value smaller than the best match value is found, the best match value and the best match region are updated until the search is completed.

6. The moving target recognition and tracking method as described in claim 1, characterized in that, Based on the best matching region of the target image, the particles in the target image are corrected to obtain the particle weights of the target image, including: Based on the best matching region of the target image, the observed values ​​of the particles are obtained, and the observation equation is: ,in For the observation equation, Let k be the particle state at time k. For the observed values, To observe noise; Obtain the observations at time k Then, the prior probability density is updated using Bayes' theorem to obtain the posterior probability density, and the posterior probability density equation is: ,in The normalization constant is This represents all observations from time 0 to time k. This represents all observations from time 0 to k-1. Let be the prior probability density. This represents the posterior probability density. The target image particles are corrected based on the posterior probability density, and the target image particle weights are updated.

7. The moving target recognition and tracking method as described in claim 1, characterized in that, The target image is resampled according to the particle weights corresponding to each particle, including: Particles with higher weights are classified as master particles, and particles with lower weights are classified as slave particles. Particles are scattered near the main particle in descending order of their weight; Retain some particles, and sprinkle fewer or no particles near the retained particles.

8. The moving target recognition and tracking method as described in claim 1, characterized in that, Based on the particle distribution after resampling of the target image, the position of the moving target in the target image is obtained, including: Based on the principal particle distribution after resampling of the target image, curve fitting is performed on the trajectory of the moving target using the least squares method. Based on the fitted trajectory of the moving target, the position of the moving target in the target image is obtained.

9. A computer-readable storage medium, characterized in that: The storage medium stores at least one instruction, which is executed by a processor to implement a moving target recognition and tracking method as described in any one of claims 1-8.

Images