Target tracking method based on sampling region adaptive search
By introducing the sampling area adaptive search strategy in the target tracking method, the problem that the KCF method cannot accurately track the target points of the jump berth in harsh environments is solved, and higher tracking accuracy and stability are achieved.
Patent Information
- Application Number
- CN202411714433.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-27
- Publication Date
- 2025-03-14
AI Technical Summary
The existing KCF-based target tracking method cannot continuously and accurately track the berth target points in harsh environments, resulting in tracking drift and reducing berth tracking performance.
The target tracking method based on adaptive search of the sampling area is adopted. By calculating the maximum feature similarity value in the feature similarity matrix, it determines whether the sampling area contains target points, and adjusts the sampling area if necessary to ensure the accuracy of the tracking.
The accurate tracking capability of the target points of the jump berth is improved, the success rate of continuous tracking is significantly improved, and the average error is reduced, and the performance and stability of target point tracking is enhanced.
Smart Images
Figure CN119625024A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of target tracking, and in particular relates to a target tracking method based on adaptive search of sampling areas. Background Art
[0002] The KCF (Kernelized Correlation Filter) tracking algorithm is an algorithm used for target tracking. The core function of the KCF algorithm is to monitor and continuously update the positioning data of a specific object in real time. The algorithm is known for its excellent processing efficiency and accuracy, and can maintain stable tracking performance even under complex and changing backgrounds and target forms. Therefore, the KCF algorithm has been widely used in the field of target tracking technology, covering multiple fields such as monitoring systems, autonomous driving technology, and drone operations. Its role is to improve the automation and accuracy of the system, so that the system can track the target more effectively and respond accordingly.
[0003] However, in harsh environments, the berth target point may jump in the AVM image (Around View Monitor) due to the maneuvering of the ego vehicle, and the existing KCF-based tracking method cannot continuously and accurately track the berth target point that has jumped, which eventually causes tracking drift and deteriorates the berth tracking performance. Summary of the invention
[0004] The object of the present invention is to provide a target tracking method based on adaptive search of sampling area.
[0005] The present invention provides a target tracking method based on sampling area adaptive search, comprising the following steps: Step 1: Set the target point as the initial target point at the initial moment; and obtain the target feature corresponding to the initial target point according to the coordinates of the initial target point.
[0006] Step 2: Build an initial classifier.
[0007] Step 3: Target point tracking 3-1. Obtain the target feature of the current moment k based on the target point k-1 at the previous moment , and calculate the feature similarity matrix based on the target feature .
[0008] 3-2. Feature similarity matrix The maximum feature similarity value max( ) and the threshold parameters set For comparison: Formula (1) Among them, max( ) is the maximum feature similarity value of k at the current moment; max( ) is the maximum feature similarity value of k-1 at the previous moment.
[0009] If the feature similarity matrix Satisfying the conditions of formula (1), the feature similarity matrix As the final feature similarity matrix of k at the current moment Otherwise, repeatedly update the sampling area and calculate the maximum feature similarity value in the feature similarity matrix , until The following relationship is satisfied: Formula (2) in, The search termination threshold is set.
[0010] The maximum feature similarity value of formula (2) The corresponding feature similarity matrix is the final feature similarity matrix at the current moment k ; 3-3. According to the final feature similarity matrix Calculate the tracking output response matrix F k , tracking output response matrix F k The expression is: Formula (3) in, is the classifier weight parameter of the previous moment k-1.
[0011] According to the response matrix F k Get the tracking target point at the current time k; use the tracking target point as the target tracking position.
[0012] Step 4: Update the target point track according to the target tracking position obtained in step 3.
[0013] Step 5: According to the track updated in step 4, train the classifier to obtain the classifier at the current time k.
[0014] Step 6: Loop steps 3 to 5 to obtain the position of the target point at different times and complete the tracking of the target point.
[0015] Preferably, in step 3, the method for updating the sampling area is: starting from the original sampling area, taking half of the side length of the original sampling area as the step length, and translating in the same direction according to the historical track of the target point corresponding to the original sampling area. In the process of updating the sampling area, if the center point of the sampling area exceeds the boundary, a sampling area is selected in the historical track as the sampling area at the current moment, and the recalculated feature similarity matrix is used as the final feature similarity matrix at the current moment. .
[0016] Preferably, the process of selecting the sampling area in the historical trajectory is: the sampling area selected in the historical trajectory is the sampling area with the smallest decrease in the maximum final feature similarity value.
[0017] Preferably, in step 4, the method for updating the target point track is divided into the following two methods: a) There are detection target points obtained using the detection network Get the track confidence of the target point ;like , then the target point track association fails, use the tracking target point to update the track information and track association failure times t i Add 1; where, is the set confidence threshold; if , calculate the detection target points obtained by the detection network And the tracking target point obtained in step 3 The Euclidean distance dis between: Formula (4) Wherein, i=1,2,...,2n; j=1,2,...,2n; n is the number of target berths.
[0018] If the Euclidean distance dis is less than or equal to the set threshold , it is considered that the detection target point is successfully associated with the target point track, and the detection target point is used to update the track information; otherwise, it is considered that the association fails, and the tracking target point is used to update the track information and the track association fails times t i Add 1, take the detected target point as the initial target point, and generate a new target point track.
[0019] b) There is no detection target point obtained using the detection network Use the tracking target point to update the track; judge the tracking result by confidence and historical track information. , and If any of the following conditions is met, the tracking result is considered inaccurate, and the number of track association failures is t i Add 1; where, is the set drift threshold.
[0020] According to the number of association failures t i Judge all updated target point tracks at the current time k; if t i > , then delete the corresponding tracking track; otherwise, keep the tracking track; where, It is the preset association failure threshold.
[0021] Preferably, the method for acquiring the target feature is: A target frame is generated with the target point coordinates as the center, and the sampling area is expanded based on the target frame. The target features corresponding to the sampling area are obtained after grayscale feature extraction.
[0022] Preferably, the method for obtaining the initial moment classifier is as follows: The classifier is trained by generating a sample set through circulant matrix sampling. The Gaussian kernel function is used to accelerate the calculation during the training process. The classifier weight parameters are obtained according to the classifier initialization formula. And the target historical feature z0, get the classifier weight parameter And the expression of the target historical feature z0 is: Formula (5) Formula (6) in, The classifier weight parameters initialized for the i-th track; is the initial target feature The corresponding feature similarity matrix; The historical target features initialized for the ith track; is the target feature extracted at the initial moment of the ith track; and λ are the preset Gaussian labels and regularization parameters respectively; i=1,2,...,2n; n is the number of berth information.
[0023] Preferably, in step 3, the method for obtaining the target tracking position is: converting the tracking output response matrix F k Convert to the coordinate system, and then convert the maximum final feature similarity value max( ) The corresponding position is taken as the target tracking position at the current moment k.
[0024] Preferably, the method for training the classifier is as follows: After obtaining the updated tracking position information at the current time k, feature extraction is performed on the sampling area corresponding to the target position obtained at the current time k, and multiple samples are generated using a circulant matrix. The Gaussian kernel function is used to accelerate classifier training, and the classifier parameters at the current time k are obtained as shown in the following formula: Formula (7) Formula (8) in, is the initial classifier weight parameter; z new is the initial historical target feature; is the target feature at time k The corresponding feature similarity matrix.
[0025] Perform weighted fusion on the initial classifier parameters of the current moment k and the classifier parameters of the previous moment k-1 to obtain the classifier parameters of the current moment k and the classifier parameters of the current moment k. and The expression is: Formula (9) Formula (10) Among them, β is the preset classifier parameter.
[0026] Preferably, the initial target point is marked by a detection network.
[0027] The present invention has the following beneficial effects: 1. The present invention introduces a sampling area adaptive search strategy, performs Gaussian kernel calculation on the initial sampling area to obtain a feature similarity matrix and a maximum response value, and adaptively adjusts the position of the sampling area based on the response value, thereby solving the problem that the traditional KCF tracking method cannot continuously and accurately track the berth target point that has jumped, and improving the accuracy of capturing the position of the berth target point that has jumped.
[0028] 2. The present invention associates the berth target points obtained from different sampling areas and processes the associated tracks, achieving significant improvements in the two key indicators of continuous tracking success rate and average error compared to the traditional KCF algorithm, thereby greatly enhancing the performance and stability of target point tracking. In actual application scenarios, this innovative technology can be seamlessly integrated into the automatic parking system in the field of unmanned driving. Faced with complex and changeable parking environments, especially when the parking target point jumps due to light changes, occlusion or other factors, the present invention can quickly adapt and accurately track, effectively avoiding misjudgment and deviation. BRIEF DESCRIPTION OF THE DRAWINGS
[0029] Figure 1 Schematic diagram of the target tracking framework of the present invention.
[0030] Figure 2 Schematic diagram of the berth target point beyond the sampling area.
[0031] Figure 3 Schematic diagram of the partial data set for secondary construction of target point jump.
[0032] Figure 4 : It is a comparison diagram of target point tracking effects; among them, (a) is the KCF target point tracking effect diagram; (b) is the target point tracking effect diagram of the present invention.
[0033] Figure 5Result diagram of target point tracking in segment 1; wherein, (a) is the target point tracking trajectory of KCF and the present invention; and (b) is the target point tracking position error of KCF and the present invention.
[0034] Figure 6 It is the target point tracking result diagram of segment 2; wherein, (a) is the target point tracking trajectory of KCF and the present invention; (b) is the target point tracking position error of KCF and the present invention. DETAILED DESCRIPTION
[0035] The present invention will be further described below in conjunction with the accompanying drawings.
[0036] like Figure 1 As shown, a target tracking method based on adaptive search in a sampling area includes the following steps: Step 1: Extract initial features Use the detection network to detect the parking information at the initial moment. The parking information is the coordinates of the two corner points on the target parking space where the vehicle needs to park. The two corner points are the upper and lower points on the outer side of the parallel parking. If the detection network outputs n parking information, there are 2n initial target points. ;in, is the position of the i-th target point at the initial moment; i=1,2,...,2n; a target frame is generated with the initial target point coordinates as the center, and the size is set to sz=[s1 s2]; where s1 is the length of the target frame; s2 is the width of the target frame. A target feature area is obtained based on the target frame expansion, which contains not only the target but also the surrounding background. The image in this area is grayed out, and the original RGB three-channel color image is converted into a single-channel grayscale image, where different grayscale values (ranging from 0 to 255) represent different brightness levels, with higher pixel values in bright areas and lower pixel values in dark areas. Normalize the pixel values of the grayscale image:
[0037] Formula (1) Among them, norm is the pixel value after normalization; I is the pixel value of different pixels before normalization; min(I) and max(I) are the minimum pixel value and maximum pixel value in the grayscale image, respectively.
[0038] After normalizing all pixels in the target feature area, the target feature corresponding to the target feature area is obtained. .
[0039] In this embodiment, the number of target points at the initial moment is the number of tracks. The detection network is a VPS-Net network.
[0040] Step 2: Get the initial classifier The sample set is expanded through the method of circular matrix sampling so as to be used for training the parameters of the classifier. During the training process, the Gaussian kernel function is used to improve the calculation efficiency, and then the weight parameters of the classifier are calculated according to the initialization formula of the classifier. And the target historical feature z0, get the classifier weight parameter And the expression of the target historical feature z0 is:
[0041] Formula (2) Formula (3) in, The classifier weight parameters initialized for the i-th track; is the initial target feature The corresponding feature similarity matrix; The historical target features initialized for the ith track; is the target feature extracted at the initial moment of the ith track; and λ are the preset Gaussian labels and regularization parameters respectively.
[0042] Step 3: Target point tracking 3-1. Calculate feature similarity Use the position information of the target point at the previous moment k-1 to establish a target frame, and set the size to sz=[s1 s2]; expand the target frame to obtain the sampling area at the current moment k, and extract the grayscale features of the sampling area to obtain the target features at the current moment k , and calculate the feature similarity matrix based on the target feature , feature similarity matrix The expression is: Formula (4) in, It is the target historical feature after k-1 updates at the last moment.
[0043] 3-2.Search sampling area adaptation like Figure 2 As shown in the figure, when the target point changes significantly, obtaining the sampling area at the current time k only based on the tracking position at the previous time k-1 will result in the target point at the current time k not being in the corresponding sampling area, and obtaining the feature similarity matrix The maximum feature similarity value max( ) will drop significantly. ) and the threshold parameters set For comparison:
[0044] Formula (5) Among them, max( ) is the maximum feature similarity value of k at the current moment; max( ) is the maximum feature similarity value of the previous moment k-1; is the set threshold parameter.
[0045] If the feature similarity matrix obtained If the above relationship is not satisfied, the feature similarity matrix As the final feature similarity matrix at the current time k Otherwise, it is considered that the sampling area does not contain the target point, and the sampling area is adjusted: starting from the original sampling area, half of the side length of the original sampling area is the step length, and the same direction translation is repeated according to the historical motion trajectory of the target point corresponding to the original sampling area, and the maximum value in the feature similarity matrix is calculated. , until The following relationship is satisfied: Formula (6) in, The search termination threshold is set.
[0046] At this time, it is considered that the target is included in the sampling area, and the maximum feature similarity value is used The corresponding feature similarity matrix is the final feature similarity matrix at the current time k .
[0047] During the translation of the sampling area, if the center point of the sampling area exceeds the boundary, the maximum final feature similarity value max( ) The sampling area with the smallest decrease is used as the sampling area at the current moment, and the recalculated feature similarity matrix is used as the final feature similarity matrix at the current moment k .
[0048] 3-3. Calculate the maximum tracking response value According to the final feature similarity matrix Calculate the tracking output response matrix F k , tracking output response matrix F k The expression is: Formula (7) in, is the classifier weight parameter of the previous moment k-1; is the dot product symbol.
[0049] Since the above steps are completed in Fourier space, the output response matrix F is tracked by inverse Fourier transform. k Convert to the coordinate system, and then convert the maximum final feature similarity value max( ) The corresponding position is taken as the tracking target point (target tracking position) at the current time k.
[0050] Step 4: Update the track 4-1. Update the target point track at the current moment according to the acquired target tracking position. Since target point tracking is based on detection results, the target point track update processing methods with and without detection results are also different. The two target point track update methods are as follows:
[0051] a) Target point track update with detection results Establish a set of detection target points based on the detection results obtained by the k detection network at the current moment ;in, is the position of the i-th detection target point at the current time k; and are the horizontal and vertical coordinates of the i-th target point k at the current moment; i=1,2,...,2n.
[0052] Establish a set of tracking target points based on the tracking results obtained by the k tracking algorithm at the current moment ;in, is the position of the jth tracking target point at the current time k; j=1,2,...,2n.
[0053] Get the track confidence of the target point ,like , then the target point track association fails, use the tracking target point to update the track information and track association failure times t i Add 1; where δ c is the set confidence threshold. , calculate the detection target point and tracking target points The Euclidean distance dis between:
[0054] Formula (8) If the Euclidean distance dis is less than or equal to the set threshold , it is considered that the detection target point is successfully associated with the target point track, and the detection target point is used to update the track information; otherwise, it is considered that the association fails, and the tracking target point is used to update the track information and the track association fails times t i Add 1. If a detected target point is found to be unsuccessfully associated with any existing target point track during the detection process, the target point should be considered as a new initial point and a new track should be created for it.
[0055] b) Target point track update without detection results Use the tracking target point to update the track; judge the tracking result by confidence and historical track information. , and If any of the following conditions is met, the tracking result is considered inaccurate, and the number of track association failures is t i Add 1; where δ e is the set drift threshold.
[0056] 4-2. According to the number of association failures t i Judge all updated target point tracks at the current time k; if t i > , then delete the corresponding tracking track; otherwise, keep the tracking track; where, It is the preset association failure threshold.
[0057] Step 5: Train the classifier The classifier is trained based on the target point track, confidence level, and number of association failures obtained in step 4 for the generation of the target point at the next moment. The classifier training method is as follows:
[0058] After obtaining the updated tracking position information at the current time k, extract features from the sampling area of the latest tracking position at the current time k. Then, generate samples through circulant matrix sampling, and use Gaussian kernel function to accelerate the training of the classifier to establish the classifier parameters at the time k. Its mathematical expression is as follows:
[0059] Formula (9) Formula (10) in, is the target feature at time k The corresponding feature similarity matrix.
[0060] Then, it is weightedly fused with the classifier parameters of the previous moment k-1 to obtain the latest classifier parameters. The update formula is as follows: Formula (11) Formula (12) Among them, β is the preset classifier parameter.
[0061] Step 6: Loop steps 3 to 5 to obtain the position of the target point at different times and complete the tracking of the target point.
[0062] Step 7: Experimental Verification In order to verify the feasibility and effectiveness of the method proposed in the present invention, experiments on target point tracking under target point jump conditions were carried out using the KCF tracking algorithm and the method provided by the present invention.
[0063] The experimental parameters set during the experiment are shown in Table 1. The experimental data are different data segments containing jump scenes based on the secondary construction of the PSV dataset. Some data segments are as follows: Figure 3 shown.
[0064] Table 1 Experimental parameter settings parameter value parameter value <![CDATA[s1]]> 40 / pixel #timg# -0.3 <![CDATA[s2]]> 40 / pixel #timg# 3 / frame β 0.02 #timg# 0.1 λ #timg# #timg# 9 / pixel <![CDATA[γ1]]> -0.7 #timg# 30 / pixel Figure 4 The experimental results of a part of the data set are shown. The experiment found that in the basic KCF tracking algorithm, the problem of tracking drift occurred at the fifth moment. This is because at the fifth moment, the position of the target point has a large jump compared to the previous moment, resulting in the failure to cover the target point in the sampling area at this moment. Therefore, when the features extracted from this sampling area are compared with the historical features of the target, an inaccurate positioning result will be obtained. The tracking method proposed in the present invention judges the possibility of containing the target in the sampling area at this moment by analyzing the maximum value of the feature similarity output for the first time at the current moment. When the target position jumps so that the target position is not in the sampling area, the maximum value of the feature similarity will decrease. The present invention will use an adaptive search strategy to adjust the position of the sampling area in order to find the area most likely to contain the target. Through this method, continuous and accurate tracking of berth target points can be achieved even in the case of a jump change in the target point.
[0065] In order to fully and deeply verify the effectiveness of the proposed method, the continuous tracking success rate and average error are introduced as key evaluation indicators to measure the continuous stability performance of the tracking algorithm and the accuracy of position prediction, respectively. By analyzing two typical clips in the experimental dataset in detail, the tracking trajectory and position error diagram (such as Figure 5 and Figure 6 ), which intuitively presents the tracking ability of the algorithm when facing significant changes in the berth target point. In addition, in order to obtain more universal conclusions, a comprehensive statistical analysis was performed on all the data segments that have undergone secondary construction. The statistical results are summarized in Table 2, which not only reflects the average performance of the algorithm on the entire dataset, but also reveals its stability and accuracy under different conditions, thereby objectively verifying the wide applicability of the proposed method and its significant results in improving tracking continuity and accuracy.
[0066] Table 2 Performance comparison of target point tracking algorithms algorithm KCF Proposed method Continuous tracking success rate 82.19% 100% Average error / pixel 19.4729 3.0101 By comparing the position trajectory diagram and position error diagram of the target point jump, it can be clearly observed that when the target point is in a stable state, the KCF tracking algorithm and the method proposed in the present invention are almost consistent in tracking effect. This phenomenon is attributed to the fact that under conventional tracking conditions, the specific optimization conditions of the present invention are not triggered, so the performance of the two is comparable. However, at the critical moment when the target point jumps, the trajectory generated by the KCF tracking algorithm has a significant position offset, resulting in a sharp increase in the error with the true target point position. In contrast, the method proposed in the present invention introduces an adaptive sampling area search strategy, which effectively responds to the sudden change of the target point and significantly improves the tracking accuracy. Furthermore, the results of statistical analysis of all data segments (as shown in Table 2) clearly show that in the data segment containing the target point jump, the continuous tracking success rate and average error of the method proposed in the present invention are significantly better than the KCF algorithm. This conclusion strongly verifies the effectiveness and superiority of the method proposed in the present invention in dealing with the target point jump scenario. Through the above experimental verification and analysis, the method proposed in the present invention shows excellent effectiveness and innovation in the data segment containing the target point jump.
Claims
1. A target tracking method based on adaptive search in a sampling area, characterized in that: The following steps are involved: Step 1: Set the position of the target point; obtain the target features according to the coordinates of the target point; Step 2: Build a classifier; Step 3: Target point tracking 3-1. Obtain the target feature of the current moment k based on the target point k-1 at the previous moment , and calculate the feature similarity matrix based on the target feature ; 3-2. Feature similarity matrix The maximum feature similarity value max( ) and the threshold parameters set For comparison: Formula (1) Among them, max( ) is the maximum feature similarity value of k at the current moment; max( ) is the maximum feature similarity value of the previous moment k-1; If the feature similarity matrix Satisfying the conditions of formula (1), the feature similarity matrix As the final feature similarity matrix of k at the current moment Otherwise, repeatedly update the sampling area and calculate the maximum feature similarity value in the feature similarity matrix , until The following relationship is satisfied: Formula (2) in, is the set search termination threshold; The maximum feature similarity value of formula (2) The corresponding feature similarity matrix is the final feature similarity matrix at the current moment k ; 3-3. According to the final feature similarity matrix Calculate the tracking output response matrix F k , tracking output response matrix F k The expression is: Formula (3) in, is the classifier weight parameter of the previous moment k-1; According to the response matrix F k Get the tracking target point at the current time k; use the tracking target point as the target tracking position; Step 4: Update the target point track according to the target tracking position obtained in step 3; Step 5: train the classifier according to the track updated in step 4 to obtain the classifier at the current time k; Step 6: Loop steps 3 to 5 to obtain the position of the target point at different times and complete the tracking of the target point.
2. The target tracking method based on sampling area adaptive search according to claim 1, characterized in that: In the step 3, the method for updating the sampling area is: starting from the original sampling area, taking half of the side length of the original sampling area as the step length, and translating in the same direction according to the historical track of the target point corresponding to the original sampling area; in the process of updating the sampling area, if the center point of the sampling area exceeds the boundary, a sampling area is selected in the historical track as the sampling area at the current moment, and the recalculated feature similarity matrix is used as the final feature similarity matrix at the current moment .
3. The target tracking method based on sampling area adaptive search according to claim 2, characterized in that: The process of selecting a sampling area in the historical trajectory is as follows: the sampling area in the selected historical trajectory is the one with the maximum final feature similarity value max( )The sampling area with the smallest decrease.
4. The target tracking method based on sampling area adaptive search according to claim 1, characterized in that: In step 4, there are two methods for updating the target point track: a) There are detection target points obtained using the detection network Get the track confidence of the target point ;like , then the target point track association fails, use the tracking target point to update the track information and track association failure times t i Add 1; where, is the set confidence threshold; if , calculate the detection target points obtained by the detection network And the tracking target point obtained in step 3 The Euclidean distance dis between: Formula (4) Where, i=1,2,...,2n; j=1,2,...,2n; n is the number of target berths; If the Euclidean distance dis is less than or equal to the set threshold , it is considered that the detection target point is successfully associated with the target point track, and the detection target point is used to update the track information; otherwise, it is considered that the association fails, and the tracking target point is used to update the track information and the track association fails times t i Add 1; if there is a detected target point that fails to be successfully associated with any existing target point track, the detected target point is used as the initial target point to generate a new target point track; b) There is no detection target point obtained using the detection network Use the tracking target point to update the track; judge the tracking result by confidence and historical track information. , and If any of the following conditions is met, the tracking result is considered inaccurate, and the number of track association failures is t i Add 1; where, is the set drift threshold; According to the number of association failures t i Judge all updated target point tracks at the current time k; if t i > , then delete the corresponding tracking track; otherwise, keep the tracking track; where, It is the preset association failure threshold.
5. The target tracking method based on sampling area adaptive search according to claim 1, characterized in that: The method for obtaining the target features is: A target frame is generated with the target point coordinates as the center, and the sampling area is expanded based on the target frame; the target features corresponding to the sampling area are obtained after grayscale features are extracted from the sampling area.
6. The target tracking method based on sampling area adaptive search according to claim 1, characterized in that: In the step 1, the construction method of the classifier at the initial moment is as follows: The classifier is trained by generating a sample set through circulant matrix sampling. The Gaussian kernel function is used to accelerate the calculation during the training process. The classifier weight parameters are obtained according to the classifier initialization formula. And the target historical feature z0, get the classifier weight parameter And the expression of the target historical feature z0 is: Formula (5) Formula (6) in, The classifier weight parameters initialized for the i-th track; is the initial target feature The corresponding feature similarity matrix; The historical target features initialized for the ith track; is the target feature extracted at the initial moment of the ith track; and λ are the preset Gaussian labels and regularization parameters respectively; i=1,2,...,2n; n is the number of berth information.
7. The target tracking method based on sampling area adaptive search according to claim 1, characterized in that: In the step 3, the method for obtaining the target tracking position is: converting the tracking output response matrix F k Convert to the coordinate system, and then convert the maximum final feature similarity value max( ) The corresponding position is taken as the target tracking position at the current moment k.
8. The target tracking method based on sampling area adaptive search according to claim 1, characterized in that: In the step 5, the method of training the classifier is as follows: After obtaining the updated tracking position information at the current time k, the target area corresponding to the target position obtained at the current time k is feature extracted, and multiple samples are generated using a circulant matrix. The Gaussian kernel function is used to accelerate the classifier training, and the classifier parameters at the current time k are obtained as shown in the following formula: Formula (7) Formula (8) in, is the initial classifier weight parameter; z new is the initial historical target feature; is the target feature at time k The corresponding feature similarity matrix.
9. The target tracking method based on sampling area adaptive search according to claim 8, characterized in that: In step 5, after the classifier is trained, the initial classifier parameters at the current moment k and the classifier parameters at the previous moment k-1 are weighted fused to obtain the classifier parameters at the current moment k. and The expression is: Formula (9) Formula (10) Among them, β is the preset classifier parameter.
10. The target tracking method based on sampling area adaptive search according to claim 1, characterized in that: The initial target point is marked by a detection network, and the detection network adopts a VPS-Net network.