Extension target feature point tracking method based on motion characteristic compensation

Abstract

The invention discloses an extended target feature point tracking method based on motion characteristic compensation, and belongs to the field of target tracking of computer vision, and the method comprises the steps: firstly constructing a matching feature optimization method based on robust feature points, reducing the number of target feature point mismatching, and achieving the estimation of optimal transformation model parameters; secondly, a tracking prediction correction method based on Kalman filtering constraint is designed, and the uncertainty of the matching tracking model is better compensated; and finally, designing a long-time and short-time fusion feature template updating strategy, fusing historical information and current target state information, enhancing the robustness of template features, effectively avoiding noise superposition, and realizing long-time-sequence stable target tracking. According to the method, the image space information and the interested feature point information are interactively interpreted in the feature matching and prediction process, so that the model of the matched features is more robust, the prediction target position error is smaller, and the long-time tracking result is more stable and accurate.

Description

Technical Field

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

[0002] As a research hotspot in the field of computer vision, target detection and tracking are widely used in intelligent video surveillance, intelligent transportation systems, intelligent visual navigation, modern military and human-computer interaction. Image target detection and tracking technology is the key technical support in video applications, and its performance directly determines the stability of the tracking application system.

[0003] The image tracking application system mainly includes several steps, including image acquisition, information processing, target detection and target tracking. Among them, the target tracking algorithm is based on video stream tracking. The complexity and uncertainty in the application scenario and the target itself affect the target tracking effect to a certain extent. For example, changes in illumination in the scene, changes in target scale, the impact of target jitter and changes when the target is occluded all pose challenges to the target tracking algorithm. Therefore, research on algorithms that can achieve stable target tracking can lay the foundation for more advanced tasks.

[0004] For tracking of extended targets, the traditional tracking method based on local features of the target extracts the underlying features of the target, such as histogram features, color features and contour features, and uses them to perform feature matching to distinguish the background and the target. However, under 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, and are more significant than other features, have a large amount of information, are relatively stable, and are relatively easier to extract. Therefore, the tracking method based on extended target feature points can achieve high-precision positioning and tracking of the target at high speed and stability, and has good applicability to changes in brightness, shape, etc. in the image. The traditional Harris corner point does not take scale changes into account, so it is impossible to obtain feature points that can be stably tracked under changing scales. Compared with the tracking algorithms based on SIFT feature points and the tracking algorithms based on KAZE feature points, the tracking algorithm based on SURF feature points has stronger scale and rotation resistance, and the calculation rate is faster than the tracking algorithm based on SIFT feature points, and the robustness of changing brightness is stronger than the tracking algorithm based on ORB feature points. Therefore, target tracking based on high-precision, high-stability and invariant SURF features and their descriptors can improve the stability and robustness of target tracking to a certain extent. However, when there is interference from similar structures and complex backgrounds in the image, local feature descriptors containing similarities are extracted, resulting in matching ambiguity in local feature matching. At this time, the position deviation caused by the wrong match is much greater than the correct match. The tracking algorithm based on SURF feature points cannot obtain accurate target prediction results, so it is necessary to remove the noise that does not conform to the optimal model and reduce the cost function. Traditional methods cannot get rid of the interference of wrong matches, and the models estimated by them deviate seriously from the true value and have low robustness.

[0005] As one of the important technologies currently used in tracking methods based on feature point matching, the random consistency algorithm can significantly improve the matching accuracy within a certain matching number range. The random sampling consistency algorithm can eliminate false matches in the image and calculate the matching matrix. However, when the number of feature points is small or there is no feature point matching, the target tracking algorithm based on feature points cannot perform at its optimal performance. Summary of the invention

[0006] In order to solve the above technical problems, the present invention provides an extended target feature point tracking method based on motion characteristic compensation. The method integrates the matching optimization based on robust feature points and the tracking prediction and correction algorithm based on Kalman filtering to eliminate the noise that does not conform to the optimal model, unbiasedly, stably and optimally predict the state information of the target, and fuses the historical information and the current target state information through the template update algorithm based on the long-short time fusion feature, so as to enhance the robustness of the template feature points and realize the stable tracking of the target within a longer time series range.

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

[0008] A method for tracking feature points of an extended target based on motion characteristic compensation comprises the following steps:

[0009] Step S110: setting method-related parameters, including target initial position, target initial length, target initial width, and matching distance threshold;

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

[0011] Step S130: applying a SURF feature point extraction algorithm to extract feature points in the target local image area to obtain SURF feature points as an initial SURF feature point template;

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

[0013] Step S150: Taking an area 1.5 times the target local area of ​​interest as the target search area, applying the SURF feature extraction method to obtain feature points of the target search area of ​​the current frame, and performing preliminary matching of the SURF feature point template and all feature points extracted at the current moment according to the fast nearest neighbor search FLANN to obtain preliminary matching results of feature point pairs;

[0014] Step S160: for the preliminary matching results of the feature point pairs, a matching feature optimization method based on spatial preprocessing is used to obtain candidate feature matching point pairs, and then a matching feature optimization method based on random consistency selection is used to obtain robust matching point pairs after matching optimization;

[0015] Step S170: obtaining an optimal homography transformation matrix based on the robust matching point pairs, and obtaining a correct predicted position of the current target;

[0016] Step S180: After obtaining the robust matching point pair, the target position of the next frame image is predicted using the Kalman tracking prediction and correction method based on motion compensation;

[0017] Step S190: updating the SURF feature point template based on the long-short time fusion feature template updating strategy.

[0018] Step S200: Output target tracking prediction results.

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

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

[0021] The present invention can solve the problem of matching ambiguity in local feature matching caused by self-similar local feature descriptors. By performing spatial preprocessing screening and random consistency screening, the number of false matches can be reduced, reliable parameter estimation results can be obtained, the robustness of target matching can be improved, and the target can be tracked accurately and stably.

[0022] When applying the Kalman tracking prediction and correction method to predict target state information, the present invention simultaneously considers the observation states of the target during stable matching and unstable matching, analyzes the target motion characteristics through matching state information and trajectory prediction information, corrects the target state under abnormal matching tracking, increases the accuracy between the tracking position and the actual position, and increases the tracking success rate.

[0023] The present invention fuses important information of the target in the dynamic region of interest of the previous template and the previous moment, so that the dynamic template has richer target characteristics, can adapt to more challenging scenes, enhance the robustness of template features, effectively avoid the superposition of noise, and achieve stable target tracking over a long time series. BRIEF DESCRIPTION OF THE DRAWINGS

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

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

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

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

[0028] FIG5 (a) is a comparison of target tracking accuracy of the present invention and the existing target tracking method for extended target tracking;

[0029] FIG5( b ) is a comparison of the success rates of extended target tracking performed by the present invention and the existing target tracking method. DETAILED DESCRIPTION

[0030] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and implementation examples.

[0031] The present invention discloses a method for tracking extended target feature points based on motion characteristic compensation, and the specific flow chart is as follows: Figure 1 As shown, the method includes:

[0032] Step S110, setting method related parameters, the related parameters include: target initial position, target initial length and target initial width, etc.

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

[0034] In one embodiment, step S120 may include: assuming that the first frame of original image data is , with the target initial position The left and right neighborhoods are 1 / 2 of the initial length of the target, and the upper and lower neighborhoods are 1 / 2 of the initial width of the target. This range is the local area of ​​interest. The image data within the local area of ​​interest is selected from the first frame of the original image data to form the target local area of ​​interest. .

[0035] Step S130: Apply the SURF feature point extraction algorithm to extract feature points in the target local image area to obtain SURF feature points as the current SURF feature point template.

[0036] The process can include:

[0037] Step a: For the local area of ​​interest of the target, use the SURF feature point extraction algorithm to generate the feature points of the local area of ​​interest of the target. The process includes: using the integral graph technology to speed up the box filter calculation speed, using different template size methods and different scale space factors of the box filter to construct image pyramids of different scales, constructing the Hessian matrix, and Image data centered , and its Hessian matrix is ​​implemented as follows:

[0038] (1)

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

[0040] The Hessian matrix is ​​used to detect extreme values ​​in the scale space and generate stable feature points in the local area of ​​interest of the target.

[0041] 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 circular neighborhood of the feature point, assign the main direction of the feature according to the Haar wavelet features in the circular neighborhood of the feature point, and finally take a square frame around the main direction of the feature point. The frame is divided into 16 sub-areas, and the Haar wavelet features of 25 pixels in the horizontal and vertical directions of each sub-area are counted to describe the feature points, and finally high-precision, high-stability and invariant SURF feature points and their feature descriptors are obtained.

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

[0043] For subsequent If it is the second frame, the second frame image data is segmented into the target local region of interest according to the target initial length and the target initial width, otherwise the target predicted position obtained in step S180 is used. The target predicted length and width are used as the segmentation area range to segment the image data and obtain the target local area of ​​interest. .

[0044] Step S150, taking an area 1.5 times the size of the target local region of interest ROI as the target search area, applying the SURF feature extraction method to obtain feature points of the target search area of ​​the current frame, and performing a preliminary match between the current SURF feature point template and all feature points extracted at the current moment based on the fast nearest neighbor search FLANN to obtain a preliminary matching result of the feature point pairs.

[0045] 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 initial original matching between the current SURF feature point template and the feature points in the target search area is realized through the nearest neighbor search algorithm based on the kd tree. The specific operation is as follows: first, a kd tree is constructed, and for the SURF feature point set and the target search area feature point set, the dimension with the highest variance in the feature descriptor is selected, and the data is divided into two parts using this dimension value, and then multiple random kd trees are established, and the dimension with the highest variance is selected. Several dimensions are randomly selected from the dimension to divide the data. When searching the random kd forest, all random kd trees share a priority queue, and finally the constructed index structure is used to obtain the nearest neighbor search matching result as the original match.

[0046] Step S160: for the preliminary matching results of feature point pairs, a matching feature optimization method based on spatial preprocessing is used to obtain candidate feature matching point pairs, and then a matching feature optimization method based on random consistency selection is used to obtain robust matching point pairs after matching optimization, such as Figure 2 shown.

[0047] The matching feature optimization method based on spatial preprocessing is based on the feature point pair The average distance of the matching of all feature points extracted in step S150 at the moment of preliminary matching result calculation As the average matching distance, and set the adaptive matching distance threshold The average matching distance is 1.5 times, and then the adaptive matching distance threshold is used to make a preliminary matching screening to obtain the preliminary screening matching results. The specific operation is to determine whether the matching distance of each pair of matching feature points exceeds the adaptive matching distance threshold. If it does not exceed the threshold, then add the candidate feature matching point pair , otherwise the matching point pairs are removed.

[0048] The matching feature optimization method based on random consistency selection first selects the template feature point set and Moment feature point set Randomly select 4 sample data from the candidate feature matching point pair set as the matching point pair set , and calculate the homography matrix based on the data in the matching point pair set , that is, the perspective transformation matrix, which contains 8 parameters of the perspective transformation matrix The calculation is as follows:

[0049] (2)

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

[0051] (3)

[0052] in, and is the first matching point pair The matching point pairs correspond to the homogeneous coordinates of the current SURF feature point template, and is the first matching point pair The homogeneous coordinates of the current frame corresponding to the matching point pairs. Based on the template feature point set and the homography matrix Computation and Moment feature point set The reprojection error is as follows:

[0053] (4)

[0054] in, The projection error is the data point whose distance is less than the distance threshold is divided into the inliers, and the matching point pairs that meet the distance threshold are judged. If the projection error is less than the error threshold, the point set that meets the optimal model is added as the current optimal inliers, otherwise it is removed; the number of iterations is updated, and if the number of iterations is greater than the maximum number of iterations, it is exited; otherwise, the number of iterations is increased by 1, and the inliers that meet the optimal model are repeatedly selected. If the number of elements in the current inliers is greater than the current optimal inliers, the current inliers that meet the optimal model are taken as the current optimal inliers, and the number of iterations is updated at the same time, and the error is calculated; finally, the robust matching point pairs are obtained, and the robust transformation matrix is ​​also obtained. , number of feature points, correct matching rate and similarity of matching results.

[0055] Step S170: Obtaining 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 pair, the Kalman tracking prediction and correction method based on motion compensation is used to predict the target position and the target length and width of the next frame image, so as to better compensate for the uncertainty of the matching tracking model. Figure 3 shown.

[0057] The motion-compensated Kalman tracking prediction and correction method uses the minimum mean square error as a 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 of the next frame image, to achieve better compensation for the uncertainty of the matching tracking model, where:

[0058] The motion-compensated Kalman tracking prediction and correction method first establishes an initial model according to the target state, including the tracking state estimation value: position, speed, etc., and then determines whether the homography transformation matrix after feature point matching exists. If it exists, the observation state is established using the current frame target state and the speed of the previous frame. Otherwise, the trajectory prediction state is used as the tracking state observation value, and the tracking state prediction and mean square error prediction are performed for the next moment. Next, the relevant model is updated, and the tracking state correction value is updated using the Kalman gain, the tracking state prediction value and the tracking state observation value, including the tracking position and speed estimation value. At the same time, the Kalman gain is used 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 motion-compensated Kalman tracking prediction correction method first starts from the first frame and and Directional target motion trajectory modeling: including tracking state , initial position ,speed , initialize the state transfer matrix , control input matrix is the identity matrix, controlling the 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 determine the homography transformation matrix after feature point matching If it exists, the predicted current frame target state is used. Establish observation state with the speed of the previous frame , otherwise the trajectory prediction state is used as the observation value to predict the tracking state and mean square error at the next moment. The prediction formula is:

[0062] (6)

[0063] is the state prediction result, and according to the error covariance matrix Estimate the error covariance matrix for the next frame To assess the uncertainty of the state:

[0064] (7)

[0065] Next, the model is updated, including the Kalman gain and the error covariance matrix. The Kalman gain handles the deviation between the predicted results and the actual observations. The update is as follows:

[0066] (8)

[0067] The target status is updated as follows:

[0068] (9)

[0069] The corrected error covariance matrix is ​​shown below:

[0070] (10)

[0071] Finally, the target tracking prediction result for the next moment is obtained, including the position ,speed :

[0072] (11)

[0073] The Kalman filter is used to dynamically adjust the estimation of the tracking results, and the update of the tracking state can achieve 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-short time fusion feature template update strategy.

[0075] When the target features change greatly, it is impossible to obtain a robust matching result. In this case, the SURF feature point, i.e., the target feature point template, is updated based on the long-short time fusion feature template update strategy. By fusing historical information and current target information, the robustness of the template feature is enhanced. The specific process is as follows: Figure 4 shown.

[0076] The long-short time fusion feature template update strategy includes: first determine whether the update conditions are met, calculate the quality score based on the matching rate and tracking distance of the matching results, and decide whether to use the long-short time fusion feature template update strategy to update the SURF feature points based on the score. Only when the score is higher than the threshold, the current SURF feature point template will be updated.

[0077] The construction method of long-term memory template is as follows: First, the matching distribution histogram of the previous template is calculated. The number of times feature points are matched and total number of matches Statistics template The frequency distribution of feature points matched :

[0078] (12)

[0079] in, is the total number of feature points. The feature points with the top 10% matching frequency are selected for homography matrix transformation and transformed to the coordinate system of the current frame to form a long-term memory template.

[0080] The construction method of short-term matching template: First, preliminarily match the feature points in the tracking result range of the previous moment with the feature points in the tracking result range of the current moment to obtain the matching results of the ROI features before and after the image data. The target characteristics of the moment and The information entropy of the target feature points at the moment is calculated for the matching point pairs at two moments, and the feature points with high entropy values ​​are selected to describe the corresponding The short-term matching template is constructed based on the feature point positions at each moment.

[0081] Finally, the strong robust feature points selected by the two templates are synthesized as the current SURF feature point template.

[0082] Step S200: output target tracking prediction results.

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

[0084] In order to verify the performance of the algorithm, the experimental verification was carried out on the Python programming platform with a memory of 16.0 GB and an 11th Gen Intel(R) Core(TM) i5-1135G7 @ 2.40GHz2.42 GHz system. The experiment used a continuous video sequence with an image size of 1280×720 as the algorithm verification data set. Figure 5 (a) is a comparison of the target tracking accuracy of different methods, and Figure 5 (b) is a comparison of the target tracking success rate of different methods. The comparison algorithms mainly include: LKOpticalFlow, MOSSE, KCF and Boosting. The experimental results show that: with the marked target position state information as a reference, the target tracking effect of the method of the present invention is the best, and the method has achieved the highest performance indicators in both accuracy and success rate. By comparing the experimental result graphs, it is found that the optical flow method, MOSSE based on correlation kernel and KCF algorithm will offset the target position when the target has a large posture change, 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 classifications, but compared with the method of the present invention, its effect is poor in accuracy and success rate indicators, and the tracking position deviation is also greater than 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 that of 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 a tracking prediction and correction method of a matching feature optimization method of robust feature points and Kalman filter constraints into an extended target tracking method, designs a matching feature optimization method based on robust feature points, eliminates mismatched feature points, improves matching accuracy, and reduces prediction errors, designs a tracking prediction and correction method based on Kalman filter constraints, corrects tracking mutations during abnormal matching, and increases tracking accuracy, and designs a long-short-term memory update module, utilizes historical information and current target state information, reduces error accumulation, achieves long-term stable tracking, and ultimately obtains better image target tracking performance.

[0085] The specific embodiments described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for tracking extended target feature points based on motion characteristic compensation, characterized in that: The steps include: Step S110: setting method-related parameters, including target initial position, target initial length, target initial width, and matching distance threshold; Step S120: for the first frame of image data, performing region segmentation on the first frame of image data according to the target initial position, the target initial length, and the target initial width to obtain a target local region of interest; Step S130: applying a SURF feature point extraction algorithm to extract feature points in the target local image area to obtain SURF feature points as SURF feature point templates; Step S140: segmenting subsequent frame image data to obtain a target local region of interest; Step S150: Taking an area 1.5 times the target local area of ​​interest as the target search area, applying the SURF feature extraction method to obtain feature points of the target search area of ​​the current frame, and performing preliminary matching of the SURF feature point template and all feature points extracted at the current moment according to the fast nearest neighbor search FLANN to obtain preliminary matching results of feature point pairs; Step S160: for the preliminary matching results of the feature point pairs, a matching feature optimization method based on spatial preprocessing is used to obtain candidate feature matching point pairs, and then a matching feature optimization method based on random consistency selection is used to obtain robust matching point pairs after matching optimization; Step S170: obtaining an optimal homography transformation matrix based on the robust matching point pairs, and obtaining a correct predicted position of the current target; Step S180: After obtaining the robust matching point pair, the target position of the next frame image is predicted using the Kalman tracking prediction and correction method based on motion compensation; Step S190: updating the SURF feature point template based on the long-short time fusion feature template updating strategy; Step S200: Output target tracking prediction results.

2. The extended target feature point tracking method based on motion characteristic compensation according to claim 1, characterized in that: The step S120 includes: Assume that the first frame of original image data is , with the target initial position The left and right neighborhood size is 1 / 2 of the target initial length, and the upper and lower neighborhood size is 1 / 2 of the target initial width. This range is the local range of interest. Image data within the local range of interest is selected from the first frame of original image data 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 circular neighborhood of the feature point, assign the main direction of the feature 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-areas, and each sub-area counts the Haar wavelet features of 25 pixels in the horizontal and vertical directions to describe the feature points, and finally obtains the SURF feature points and their feature descriptors as the initial SURF feature point templates.

4. The extended target feature point tracking method based on motion characteristic compensation according to claim 3 is characterized in that: 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 extended target feature point tracking method based on motion characteristic compensation according to claim 1, characterized in that: 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 of 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 nearest neighbor search algorithm based on kd tree includes: firstly, constructing kd tree, selecting the dimension with the highest variance in feature descriptor for SURF feature point set and target search area feature point set, dividing the data into two parts by using this dimension value, and then establishing multiple random kd trees, starting from the one with the highest variance. Several dimensions are randomly selected from the dimension to divide the data. When searching the random kd forest, all random kd trees share a priority queue, and finally the constructed index structure is used to obtain the nearest neighbor search matching results.

8. The extended target feature point tracking method based on motion characteristic compensation according to claim 1, characterized in that: The step S160 includes: The matching feature optimization method based on spatial preprocessing calculates the average matching distance of all feature points extracted in step S150 according to the preliminary matching results of feature point pairs, and sets the adaptive matching distance threshold to 1.5 times the average matching distance. Then, the matching results are preliminarily screened according to the adaptive matching distance threshold: it is determined 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, it is added to the candidate feature matching point pair set, otherwise it is eliminated. The matching feature optimization method based on random consistency selection first randomly selects 4 sample data from the candidate feature matching point pair set as the matching point pair set, and calculates the homography transformation matrix based on the data in the matching point pair set; calculates the projection errors of all feature matching point pairs in the candidate feature matching point pair set in the homography transformation matrix, and divides the data points whose distance is less than the distance threshold into inliers. The matching point pairs that meet the distance threshold are judged. If their projection error is less than the error threshold, they are added to the inliers set, otherwise they are removed, and the optimal inliers set is updated, and the number of iterations is updated at the same time; if the number of iterations is greater than the maximum number of iterations, exit; otherwise, the number of iterations is increased by 1, and the selection and error calculation are repeated to finally obtain robust matching point pairs.

9. The extended target feature point tracking method based on motion characteristic compensation according to claim 1, characterized in that: 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 the calculation of the estimated value of the next frame position and the prediction of the target position of the next frame image, wherein: The motion-compensated Kalman tracking prediction and correction method first establishes an initial tracking model according to the target state, and then determines whether the homography transformation matrix after feature point matching exists. If it exists, the observation state is established using the current frame target state and the speed of the previous frame. Otherwise, the trajectory prediction state is used as the tracking state observation value, and the tracking state and mean square error are predicted for the next moment. Next, the relevant update of the tracking model is performed, and the tracking state correction value is updated using the Kalman gain, the tracking state prediction value and the tracking state observation value. At the same time, the mean square error of the result is corrected using the Kalman gain to realize dynamic adjustment estimation of the tracking result.

10. The extended target feature point tracking method based on motion characteristic compensation according to claim 1, characterized in that: The step S190 includes: constructing a long-term memory template and a short-term matching template simultaneously based on a long-term and short-term fusion feature template update strategy, wherein: The long-short time fusion feature template update strategy includes: firstly judging whether the update condition is met, calculating the quality score according to the matching rate and tracking distance of the matching result, and deciding whether to use the long-short time fusion feature template update strategy to update the SURF feature point template according to the score; The method for constructing the long-term memory template comprises: firstly calculating the matching distribution histogram of the previous template, counting the frequency distribution of each feature point in the template, selecting the feature points with the matching distribution frequency in the top 10% to perform homography matrix transformation, transforming to the coordinate system of the current frame, and constructing the long-term memory template; The method for constructing the short-term matching template comprises: firstly preliminarily matching the feature points within the tracking result range of the previous moment with the feature points within the tracking result range of the current moment, obtaining the matching results of the ROI features before and after the image data, and then The target characteristics of the moment and The information entropy of the target feature points at the moment is calculated for the matching point pairs at two moments, and the feature points with high entropy values ​​are selected to describe the corresponding Construct a short-term matching template based on the feature point position at each moment; Finally, the strong feature points selected by the two templates are synthesized as the new SURF feature point template.

Images