Extended Target Feature Point Tracking Method Based on Motion Characteristic Compensation

Abstract

The present invention discloses an extended target feature point tracking method based on motion characteristic compensation, belonging to the field of target tracking in computer vision. The method includes: First, a matching feature optimization method based on robust feature points is constructed to reduce the number of incorrect matches of target feature points and realize the estimation of the optimal transformation model parameters; Second, a tracking prediction correction method based on Kalman filter constraint is designed to better compensate for the uncertainty of the matching tracking model; Finally, a long-short-term fusion feature template update strategy is designed to fuse historical information and current target state information, enhance the robustness of the template features, effectively avoid the superposition of noise, and can realize stable tracking of the target in a long time series. In the feature matching and prediction process of the present invention, the image space information and the information of the feature points of interest are interactively interpreted, making the model of the matching features more robust, the prediction target position error smaller, and the long-term tracking result more stable and accurate.

Description

Technical Field

[0001] The present invention belongs to the field of object tracking in computer vision, and specifically relates to an extended object feature point tracking method based on motion characteristic compensation, which can be used to solve the problem of stable tracking of extended objects. Background Art

[0002] Object detection and tracking, as a research hotspot in the field of computer vision, is widely applied in fields such as intelligent video surveillance, intelligent transportation systems, intelligent visual navigation, modern military, and human-computer interaction. As a key technical support in video applications, the performance of image object detection and tracking technology directly determines the stability of the tracking application system.

[0003] An image tracking application system mainly includes several steps such as image acquisition, information processing, object detection, and object tracking. Among them, the object tracking algorithm is based on video stream tracking. The complexity and uncertainty existing in the application scenario and the object itself cause the effect of object tracking to be affected to a certain extent. For example, changes in illumination in the scenario, changes in object scale, the influence of object jitter, and changes when the object is occluded all pose challenges to the effect of the object tracking algorithm. Therefore, researching algorithms that can achieve stable object tracking can lay a foundation for more advanced tasks.

[0004] For tracking extended targets, traditional tracking methods based on local target features extract underlying features of the target, such as histogram features, color features, and contour features, and use these to perform feature matching to distinguish between background and target. However, in complex scene conditions, the performance of the feature matching algorithm will cause local or global changes in the matching scene. As the basis for higher-level signal representation, point features have strong invariance, stability, and resolution. They are more significant than other features, contain more information, are relatively stable, and are relatively easier to extract. Therefore, tracking methods based on extended target feature points can achieve high-precision positioning and tracking of targets at high speed and stability, and are well adapted to changes in brightness, shape, and other aspects of the image. Traditional Harris corner points do not consider scale changes, and therefore cannot obtain feature points that can be stably tracked under varying scales. Compared to tracking algorithms based on SIFT and KAZE feature points, tracking algorithms based on SURF feature points are more resistant to scale and rotation, have a faster computational speed, and are more robust to varying brightness than tracking algorithms based on ORB feature points. Therefore, target tracking based on the high-precision, highly stable, and invariant SURF features and their descriptors can improve the stability and robustness of target tracking to a certain extent. However, when similar structures and complex backgrounds interfere with the image, extracting local feature descriptors containing similar features can lead to matching ambiguity in local feature matching. In this case, the position deviation caused by mismatches is much greater than that of correct matches, and tracking algorithms based on SURF feature points cannot accurately predict the target. Therefore, it is necessary to remove noise that does not conform to the optimal model and reduce the cost function. Traditional methods cannot escape the interference of mismatches, and the models estimated using them deviate significantly from the true value, resulting in low robustness.

[0005] The randomized consensus algorithm, a key technique currently used in tracking methods based on feature point matching, can significantly improve matching accuracy within a certain range of matching numbers. It uses the randomized consensus algorithm to eliminate false matches in an image and calculate the matching matrix. However, feature point-based object tracking algorithms cannot achieve optimal performance when the number of feature points is small or no feature points are matched. Summary of the Invention

[0006] To address the above technical issues, the present invention provides an extended target feature point tracking method based on motion characteristic compensation. This method integrates robust feature point matching optimization with a Kalman filter-based tracking prediction and correction algorithm to eliminate noise that does not conform to the optimal model. This method predicts the target state information in an unbiased, stable, and optimal manner. Furthermore, a template update algorithm based on long- and short-term fusion features fuses historical information with current target state information, enhancing the robustness of template feature points and achieving stable target tracking over a longer time range.

[0007] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0008] An extended target feature point tracking method based on motion characteristic compensation, comprising the following steps:

[0009] Step S110: Set the parameters related to the method, including the initial position of the target, the initial length of the target, the initial width of the target, and the matching distance threshold;

[0010] Step S120: For the first-frame image data, perform region segmentation on the first-frame image data according to the initial position of the target, the initial length of the target, and the initial width of the target to obtain the target local region of interest;

[0011] Step S130: Apply the SURF feature point extraction algorithm to extract the feature points in the target local image region to obtain the SURF feature points, which serve as the initial SURF feature point template;

[0012] Step S140: Segment the subsequent frame image data to obtain the target local region of interest;

[0013] Step S150: Take a region 1.5 times the target local region of interest as the target search region, apply the SURF feature extraction method to obtain the feature points in the target search region of the current frame, and perform a preliminary match on the SURF feature point template and all the feature points extracted at the current moment according to the fast nearest neighbor search FLANN to obtain the preliminary matching result of the feature point pairs;

[0014] Step S160: For the preliminary matching result of the feature point pairs, use the matching feature optimization method based on spatial preprocessing to obtain the candidate feature matching point pairs, and then use the matching feature optimization method based on random consensus selection to obtain the robust matching point pairs after matching optimization;

[0015] Step S170: Obtain the optimal homography transformation matrix based on the robust matching point pairs, and obtain the correct prediction position of the current target;

[0016] Step S180: After obtaining the robust matching point pairs, use the Kalman tracking prediction correction method based on motion compensation to realize the prediction of the target position in the next frame image;

[0017] Step S190: Update the SURF feature point template based on the long-term and short-term fusion feature template update strategy.

[0018] Step S200: Output the target tracking prediction result.

[0019] The beneficial effects of the present invention are as follows:

[0020] The object tracking method based on Kalman filtering uses the minimum mean square error as the criterion, establishes an equation through the observed data and the model prediction data, and then corrects the observed target position. It can not only effectively handle the occlusion problem to a certain extent, but also better compensate for the uncertainty of the model, providing more accurate and stable target state information for the model.

[0021] The present invention can solve the problem of matching ambiguity in local feature matching caused by self-similar local feature descriptors. Through spatial preprocessing screening and random consensus screening, it can reduce the number of incorrect matches, obtain reliable parameter estimation results, improve the robustness of target matching, and accurately and stably track the target.

[0022] When the present invention predicts the target state information using the Kalman tracking prediction correction method, it simultaneously considers the observed states in both stable and unstable target matches. By analyzing the matching state information and the trajectory prediction information, it corrects the target state under abnormal matching tracking, increasing the accuracy between the tracking position and the true position and increasing the success rate of tracking.

[0023] The present invention fuses the important information of the target in the previous template and the dynamic region of interest at the previous moment, making the dynamic template have richer target characteristics, being able to adapt to more challenging scenarios, enhancing the robustness of the template features, effectively avoiding the superposition of noise, and achieving long-term stable tracking of the target. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 is the flowchart of the extended target feature point tracking method based on motion characteristic compensation of the present invention;

[0025] Figure 2 is the schematic diagram of the matching optimization algorithm based on robust feature points of the present invention;

[0026] Figure 3 is the schematic diagram of the tracking prediction correction algorithm based on Kalman filtering of the present invention;

[0027] Figure 4 is the schematic diagram of the template update algorithm based on long and short-term fusion features;

[0028] Figure 5(a) shows the comparison of target tracking accuracy for extended target tracking between the present invention and existing target tracking methods;

[0029] Figure 5(b) shows the comparison of success rates for extended target tracking between the present invention and existing target tracking methods. DETAILED DESCRIPTION OF THE INVENTION

[0030] The technical solutions of the present invention will be described in detail below in conjunction with the accompanying drawings and embodiments.

[0031] The present invention discloses an extended target feature point tracking method based on motion characteristic compensation. The specific flowchart is as shown in Figure 1 and the method includes:

[0032] Step S110, set parameters related to the method. The related parameters include: the initial position of the target, the initial length of the target, the initial width of the target, etc.

[0033] Step S120, for the first-frame image data, perform region segmentation on the first-frame image data according to the initial position of the target, the initial length of the target, and the initial width of the target to obtain the target local region of interest.

[0034] In an embodiment, step S120 may include: assuming that the first-frame original image data is , with the initial position of the target as the center, the left and right neighborhood sizes are 1 / 2 of the initial length of the target, and the upper and lower neighborhood sizes are 1 / 2 of the initial width of the target. This range is the local region of interest range. Select the image data within this local region of interest range from the first-frame original image data to form the target local region of interest .

[0035] Step S130, apply the SURF feature point extraction algorithm to extract the feature points in the target local image region to obtain the SURF feature points, which are used as the current SURF feature point template.

[0036] This process may include:

[0037] Step a, for the target local region of interest, use the SURF feature point extraction algorithm to generate the feature points of the target local region of interest. This process includes: accelerating the calculation speed of the box filter through the integral image technique, constructing different-scale image pyramids by using different template sizes of the box filter and different scale space factors, constructing the Hessian matrix, and for the image data centered on , its Hessian matrix is implemented as follows:

[0038] (1)

[0039] where is the first-order partial derivative operation, and is the second-order partial derivative operation.

[0040] Perform extreme value detection in the scale space through the Hessian matrix to generate stable feature points in the target local region of interest.

[0041] Step b: Apply non-maximum suppression for feature point filtering and 3D linear interpolation algorithm for precise positioning. Take the circular area with a fixed radius centered on the feature point as the circular neighborhood of the feature point. Assign the principal direction of the feature based on the Haar wavelet features within the circular neighborhood of the feature point. Finally, take a square box around the principal direction of the feature point. The box is divided into 16 sub-regions, and the horizontal and vertical Haar wavelet features of 25 pixels are counted in each sub-region to describe the feature point, and finally obtain high-precision, high-stability and invariant SURF feature points and their feature descriptors.

[0042] Step S140: Segment the subsequent frame image data to obtain the target local region of interest.

[0043] For the subsequent frame of image data, if it is the second frame, segment the second frame of image data with the target initial length and target initial width to obtain the target local region of interest. Otherwise, use the target predicted position obtained in step S180 as the center position, and the target predicted length and width as the segmentation region range to segment the image data to obtain the target local region of interest .

[0044] Step S150: Take 1.5 times the area of the target local region of interest ROI as the target search region. Apply the SURF feature extraction method to obtain the feature points in the target search region of the current frame. Based on the fast nearest neighbor search FLANN, perform a preliminary match between the current SURF feature point template and all the feature points extracted at the current moment to obtain the preliminary match result of the feature point pairs.

[0045] Based on the target search region, apply the SURF feature point extraction algorithm to extract the feature points in the target search region. Implement the preliminary raw match between the current SURF feature point template and the feature points in the target search region through the nearest neighbor search algorithm based on the k-d tree. The specific operation is as follows: First, construct a k-d tree for the SURF feature point set and the target search region feature point set. Select the dimension with the highest variance in the feature descriptor, and use the numerical value of this dimension to divide the data into two parts. Then, build multiple random k-d trees, randomly select several dimensions from the dimensions with the highest variance to divide the data. When searching the random k-d forest, all random k-d trees share a priority queue. Finally, use the constructed index structure to obtain the nearest neighbor search match result as the raw match.

[0046] Step S160: For the preliminary match result of the feature point pairs, use the matching feature optimization method based on spatial preprocessing to obtain the candidate feature matching point pairs, and then use the matching feature optimization method based on random consensus selection to obtain the robust matching point pairs after matching optimization, as Figure 2 shown.

[0047] The matching feature optimization method based on spatial preprocessing is based on the feature point pairs Calculate the average distance of the matches of all the feature points extracted in step S150 of the preliminary matching result at a certain moment As the average matching distance, and set an adaptive matching distance threshold To be 1.5 times the average matching distance, and then make a judgment according to the adaptive matching distance threshold, perform preliminary screening of the matching, and obtain the preliminary screening matching result. The specific operation is to judge whether the matching distance of each pair of matching feature point pairs exceeds the adaptive matching distance threshold If it does not exceed the threshold, add it to the candidate feature matching point pairs Otherwise, eliminate the matching point pairs

[0048] The matching feature optimization method based on random consensus selection first selects 4 sample data randomly from the template feature point set and The feature point set at a certain moment Get 4 sample data randomly from the obtained candidate feature matching point pairs as the matching point pair set And calculate the homography matrix according to the data in the matching point pair set That is, the perspective transformation matrix, the perspective transformation matrix containing 8 parameters The calculation is as follows

[0049] (2)

[0050] Among them, to recover the 8 parameters in the perspective transformation matrix, the transformation process of the randomly selected 4 sample data is as follows

[0051] (3)

[0052] Among them and Are the homogeneous coordinates of the th matching point pair in the matching point pair set corresponding to the current SURF feature point template and Are the homogeneous coordinates of the current frame corresponding to the th matching point pair in the matching point pair set. Based on the template feature point set and the homography matrix Calculate the reprojection error with the Feature point set at a certain moment According to the projection mapping result, there is

[0053] (4)

[0054] Among them For the projection error, data points with a distance less than the distance threshold are classified as inliers. Matching point pairs that satisfy the distance threshold are judged. If the projection error is less than the error threshold, they are added to the point set that conforms to the optimal model as the current optimal inlier set; otherwise, they are excluded. Update the iteration count. Judge if the iteration count is greater than the maximum iteration count, then exit; otherwise, increment the iteration count by 1 and continue to repeat the selection of the inlier set that conforms to the optimal model. If the number of elements in the current inlier set is greater than the current optimal inlier set, then use the current inlier set that conforms to the optimal model as the current optimal inlier set, and at the same time update the iteration count and calculate the error. Finally, obtain the robust matching point pairs and also obtain the robust transformation matrix , the number of feature points, the correct matching rate, and the similarity of the matching results.

[0055] Step S170: Obtain the optimal homography transformation matrix based on the robust matching point pairs , apply the optimal homography transformation matrix to transform the target position of the previous frame to obtain the correct matching position of the current target.

[0056] Step S180: After obtaining the robust matching point pairs, use the Kalman tracking prediction and correction method based on motion compensation to realize the prediction of the target position in the next frame image and the prediction of the target length and width, and better compensate for the uncertainty of the matching tracking model, as Figure 3 shown.

[0057] The Kalman tracking prediction and correction method based on motion compensation uses the minimum mean square error as the criterion to establish an equation to correct the observed target position, including the calculation of the estimated value of the next frame position and the prediction of the target position in the next frame image, and better compensates for the uncertainty of the matching tracking model, where:

[0058] The Kalman tracking prediction and correction method based on motion compensation first establishes an initial model according to the target state, including the estimated values of the tracking state: position, speed, etc., and then judges whether the homography transformation matrix after feature point matching exists. If it exists, use the current frame target state and the speed of the previous frame to establish the observation state; otherwise, use the trajectory prediction state as the tracking state observation value to predict the tracking state and the mean square error for the next moment. Next, perform the relevant update of the model, and use the Kalman gain, the tracking state prediction value, and the tracking state observation value to update the tracking state correction value, including the estimated values of the tracking position and speed. At the same time, use the Kalman gain to correct the result mean square error to realize the dynamic adjustment estimation of the tracking result. The update of the tracking state realizes the unbiased, stable, and optimal prediction of the target position and speed.

[0059] The Kalman tracking prediction and correction method based on motion compensation first starts from the first frame and separately establishes models for the target motion trajectories in and directions: including the tracking state , the initial position , the velocity , initialize the state transition matrix , the control input matrix is the identity matrix, the control input , the observation noise covariance matrix is the identity matrix, the process noise covariance matrix is 0.1 times the identity matrix, the state estimation error covariance matrix is the identity matrix, where:

[0060] (5)

[0061] Then judge the homography transformation matrix after feature point matching , if it exists, use the predicted current frame target state and the velocity of the previous frame to establish the observation state , otherwise, apply the trajectory prediction state as the observation value to predict the tracking state and the mean square error for the next moment. The prediction formula is:

[0062] (6)

[0063] is the state prediction result, and at the same time, according to the error covariance matrix estimate the error covariance matrix of the next frame to evaluate the uncertainty of the state:

[0064] (7)

[0065] Next, perform the relevant updates of the model, including the Kalman gain and the error covariance matrix, where the Kalman gain realizes the processing of the deviation between the prediction result and the actual observation, and the Kalman gain is updated as follows:

[0066] (8)

[0067] The update of the target state is as follows:

[0068] (9)

[0069] The corrected error covariance matrix is as follows:

[0070] (10)

[0071] Finally, obtain the target tracking prediction result for the next moment, including the position , the velocity :

[0072] (11)

[0073] The dynamic adjustment estimation of the tracking result is realized through Kalman filtering, and the update of the tracking state realizes unbiased, stable and optimal prediction of the position and speed of the target.

[0074] Step S190: Update the current SURF feature point template based on the long-term and short-term fusion feature template update strategy.

[0075] When the target features change greatly, a robust matching result cannot be obtained. At this time, the long-term and short-term fusion feature template update strategy is applied to update the SURF feature points, that is, the target feature point template. By fusing historical information and current target information, the robustness of the template features is enhanced. The specific process is as Figure 4 shown.

[0076] The long-term and short-term fusion feature template update strategy includes: first, judge whether the update condition is met, calculate the quality score according to the matching rate and tracking distance of the matching result, and decide whether to use the long-term and short-term fusion feature template update strategy to update the SURF feature points according to the score. Only when the score is higher than the threshold will the current SURF feature point template be updated.

[0077] The construction method of the long-term memory template: First, calculate the matching distribution histogram of the previous template, and based on the number of times the th feature point is matched and the total number of matches to statistically calculate the frequency distribution of the th feature point in the template

[0078] (12)

[0079] Among them, is the total number of feature points. Select the feature points with the top 10% matching frequencies for homography matrix transformation, transform them to the coordinate system of the current frame, and construct the long-term memory template.

[0080] The construction method of the short-term matching template: First, preliminarily match the feature points within the tracking result range of the previous moment and the feature points within the tracking result range of the current moment to obtain the matching result of the ROI features before and after the image data. On this basis, based on the target features at the th moment and the target feature points at the

[0081] Finally, synthesize the strongly robust feature points selected by the two templates as the current SURF feature point template.

[0082] Step S200: Output the target tracking prediction result.

[0083] According to the above steps, obtain the high-precision tracking result of the extended target and output this result.

[0084] To verify the performance of the algorithm, experiments were carried out on a Python programming platform with a system of 11th Gen Intel(R) Core(TM) i5-1135G7 @ 2.40GHz 2.42 GHz and a memory of 16.0 GB. The experiment used a continuous video sequence with an image size of 1280×720 as the algorithm verification dataset. Figure 5(a) shows the comparison of the target tracking accuracy of different methods, and Figure 5(b) shows the comparison of the success rate of target tracking of different methods. The comparison algorithms mainly include: LKOpticalFlow, MOSSE, KCF, and Boosting. The experimental results show that: taking the labeled target position state information as a reference, the method of the present invention has the best tracking effect for the target. The present method has obtained the highest performance indicators in both accuracy and success rate. Through the comparison of the experimental result graphs, it is found that the optical flow method, the correlation kernel-based MOSSE and KCF algorithms have offsets in the target position during tracking when the target has a large pose transformation, and the offset distance is very large. The Boosting method can improve the tracking effect of the target position to a certain extent by integrating multiple weak classifiers. However, compared with the method of the present invention, its effect is poor in terms of accuracy and success rate indicators, and the tracking position deviation is also greater than that of the method of the present invention. Taking the IoU value of 0.8 as the success rate threshold, the tracking success rate of the present invention reaches 100%, while the Boosting method is only 98.97%. The tracking success rate of the present invention is improved by 1.03% compared with the Boosting method. The method of the present invention introduces the tracking prediction correction method constrained by the Kalman filter, which is an optimization method for the matching features of robust feature points, into the extended target tracking method. It designs an optimization method for the matching features of robust feature points to eliminate mis-matched feature points, improve the matching accuracy, and reduce the prediction error. It designs a tracking prediction correction method constrained by the Kalman filter to correct the tracking mutation during abnormal matching and increase the tracking accuracy. At the same time, it designs a long short-term memory update module to use historical information and the current target state information to reduce the accumulation of errors and achieve long-term stable tracking, and finally obtains better image target tracking performance.

[0085] The specific embodiments described above further elaborate on the objectives, technical solutions, and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. An extended target feature point tracking method based on motion characteristic compensation, characterized in that It includes the following steps: Step S110: Set the method-related parameters, including the target initial position, target initial length, target initial width, and matching distance threshold; Step S120: For the first-frame image data, perform region segmentation on the first-frame image data according to the target initial position, target initial length, and target initial width to obtain the target local region of interest; Step S130: Apply the SURF feature point extraction algorithm to extract the feature points in the target local image region to obtain the SURF feature points, which serve as the SURF feature point template; Step S140: Segment the subsequent frame image data to obtain the target local region of interest; Step S150: Use a 1.5-fold region of the target local region of interest as the target search region, apply the SURF feature extraction method to obtain the feature points in the target search region of the current frame, and perform a preliminary match on the SURF feature point template and all the feature points extracted at the current moment based on the fast nearest neighbor search FLANN to obtain the preliminary match result of the feature point pairs; Step S160: For the preliminary match result of the feature point pairs, use the matching feature optimization method based on spatial preprocessing to obtain the candidate feature matching point pairs, and then use the matching feature optimization method based on random consensus selection to obtain the robust matching point pairs after optimization. The matching feature optimization method based on spatial preprocessing calculates the average matching distance of all the feature points extracted in Step S150 according to the preliminary match result of the feature point pairs, sets the adaptive matching distance threshold to 1.5 times the average matching distance, and then preliminarily filters the match result: determine whether the matching distance of each pair of matching feature point pairs exceeds the adaptive matching distance threshold. If it does not exceed the threshold, add it to the candidate feature matching point pair set; otherwise, eliminate it; Step S170: Obtain the optimal homography transformation matrix based on the robust matching point pairs and obtain the correct prediction position of the current target; Step S180: After obtaining the robust matching point pairs, use the Kalman tracking prediction correction method based on motion compensation to realize the prediction of the target position in the next frame of image; Step S190: Update the SURF feature point template based on the long-term and short-term fusion feature template update strategy. The long-term and short-term fusion feature template update strategy includes: first, determine whether it meets the update condition, calculate the quality score according to the matching rate and tracking distance of the matching result, and decide whether to use the long-term and short-term fusion feature template update strategy to update the SURF feature point template according to the score; Step S190 includes: simultaneously constructing a long-term memory template and a short-term matching template based on the long-term and short-term fusion feature template update strategy, where: The construction method of the long-term memory template includes: first, calculate the matching distribution histogram of the previous template, count the frequency distribution of each feature point in the template being matched, select the feature points with the top 10% matching distribution frequency for homography matrix transformation, and transform them to the coordinate system of the current frame to form the long-term memory template; The method for constructing the short-term matching template includes: First, initially match the feature points within the range of the tracking result at the previous moment and the feature points within the range of the tracking result at the current moment to obtain the matching result of the ROI features before and after the image data. Based on this, the target features at the moment and the feature points of the target at the moment are used to calculate the information entropy of the matching point pairs at the two moments, and the feature point description with a high entropy value is selected to construct the short-term matching template corresponding to the feature point position at the Finally, synthesize the strong feature points selected by the two-way templates as the new SURF feature point template; Step S200: Output the target tracking prediction result.

2. The extended target feature point tracking method based on motion characteristic compensation according to claim 1, wherein Step S120 includes: Assume that the original image data of the first frame is , centered at the target initial position . The left and right neighborhood sizes are 1 / 2 of the target initial length, and the up and down neighborhood sizes are 1 / 2 of the target initial width. This range is the local region of interest. Select the image data within this local region of interest from the original image data of the first frame to form the target local region of interest.

3. The extended target feature point tracking method based on motion characteristic compensation according to claim 1, characterized in that The step S130 includes: Step a: for the target local area of interest, using the SURF feature point extraction algorithm to generate feature points of the target local area of interest; Step b, apply non-maximum suppression to filter feature points and three-dimensional linear interpolation algorithm for precise positioning, take the feature point as the center and the circular area with a fixed radius as the feature point circular neighborhood, assign the feature main direction according to the Haar wavelet feature in the circular neighborhood of the feature point, and finally take a square frame around the main direction of the feature point. The square frame is divided into 16 sub-regions, and the Haar wavelet features of 25 pixels in the horizontal and vertical directions of each sub-region are counted to describe the feature point, and finally the SURF feature point and its feature descriptor are obtained as the initial SURF feature point template.

4. The extended target feature point tracking method based on motion characteristic compensation according to claim 3, wherein Step a includes: accelerating the box filter calculation speed through integral graph technology, constructing image pyramids of different scales using different template sizes and different scale space factors of the box filter, constructing a Hessian matrix, realizing extreme value detection in scale space through the Hessian matrix, and generating stable feature points in the local area of interest of the target.

5. The method for tracking extended target feature points based on motion characteristic compensation according to claim 1, wherein Step S140 includes: for the subsequent frame image data, if it is the second frame, the second frame image data is segmented into a target local region of interest using the target initial length and the target initial width; otherwise, the image data is segmented using the target predicted position as the center position and the target predicted length and the target predicted width as the segmentation area range to obtain the target local region of interest.

6. The extended target feature point tracking method based on motion characteristic compensation according to claim 1, characterized in that The step S150 includes: Based on the target search area, the SURF feature point extraction algorithm is applied to extract the feature points in the target search area, and the preliminary matching between the SURF feature point template and the feature points in the target search area is achieved through the nearest neighbor search algorithm based on the kd tree.

7. The extended target feature point tracking method based on motion characteristic compensation according to claim 6, characterized in that The k-d tree-based nearest neighbor search algorithm includes: First, construct a k-d tree. For the SURF feature point set and the target search area feature point set, select the dimension with the highest variance in the feature descriptor, use the numerical value of this dimension to divide the data into two parts, and then build multiple random k-d trees. Randomly select several dimensions from the dimensions with the highest variance to divide the data. When searching the random k-d forest, all random k-d trees share a priority queue. Finally, use the constructed index structure to obtain the nearest neighbor search matching result.

8. The method for tracking extended target feature points based on motion characteristic compensation according to claim 1, wherein In step S160, the matching feature optimization method based on random consistency selection first randomly extracts 4 sample data from the candidate feature matching point pair set as a matching point pair set, and calculates a homography transformation matrix based on the data in the matching point pair set; Calculate the projection errors of all feature matching point pairs in the candidate feature matching point pair set in the homography transformation matrix. Data points with distances less than the distance threshold are classified as inliers. Determine the matching point pairs that meet the distance threshold. If their projection errors are less than the error threshold, add them to the inlier set; otherwise, remove them. Update the optimal inlier set and the number of iterations at the same time. If the number of iterations is greater than the maximum number of iterations, exit. Otherwise, add 1 to the number of iterations, and continue to repeatedly select and calculate the errors to finally obtain robust matching point pairs.

9. The method for tracking extended target feature points based on motion characteristic compensation according to claim 1, wherein The motion-compensated Kalman tracking prediction and correction method in step S180 uses the minimum mean square error as a criterion to establish an equation to correct the observed target position, including calculating an estimated value of the next frame position and predicting the target position of the next frame image, wherein: The Kalman tracking prediction correction method based on motion compensation first establishes an initial tracking model according to the target state, and then determines whether the homography transformation matrix exists after feature point matching. If it exists, an observation state is established using the target state of the current frame and the velocity of the previous frame. Otherwise, the trajectory prediction state is used as the tracking state observation value to predict the tracking state and the mean square error for the next moment. Next, relevant updates of the tracking model are performed, and the Kalman gain, the tracking state prediction value, and the tracking state observation value are used to update the tracking state correction value. At the same time, the mean square error of the correction result is corrected using the Kalman gain to achieve dynamic adjustment and estimation of the tracking result.

Images