Target tracking method and system based on multilayer features, storage medium and equipment
By employing a target tracking method that combines multi-layer feature fusion and adaptive Kalman filtering, the robustness problem of traditional methods in complex environments is solved, achieving high-precision target tracking.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XIDIAN UNIV
- Filing Date
- 2025-12-15
- Publication Date
- 2026-04-14
AI Technical Summary
Traditional target tracking methods rely on manually designed features, lack robustness, struggle to cope with complex environmental changes and situations such as occlusion and deformation, and lack effective feedback mechanisms, leading to tracking drift.
A multi-layer feature-based target tracking method is adopted. The VGG-19 network is used to extract depth features at different levels. The feature quality is measured by the peak sidelobe ratio (PSR) and adaptive fusion is performed. Target localization and scale estimation are performed based on confidence analysis, and the method is updated by combining adaptive Kalman filtering.
It significantly improves the accuracy and robustness of target tracking algorithms, enabling accurate target tracking in complex scenarios and avoiding tracking drift.
Smart Images
Figure CN121861069A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of target tracking technology, and in particular to target tracking methods, systems, storage media, and devices based on multi-layer features. Background Technology
[0002] In the field of target tracking technology, traditional tracking methods heavily rely on hand-designed features (such as HOG, CN, color features, etc.) to describe the target's appearance. These hand-designed features may perform well in simple scenarios, but their representational capabilities are limited, and they lack robustness to complex environmental changes, such as changes in lighting, target deformation, and rapid movement. Furthermore, when fusing multiple features, traditional methods often employ simple weighted averaging or fixed fusion strategies, failing to fully consider the quality and reliability of different feature channels themselves, resulting in the feature representation capabilities not being optimally utilized.
[0003] In practical applications, targets often encounter complex situations such as occlusion, deformation, and interference from similar objects. When these situations occur, the tracker's localization results become unreliable. However, traditional correlation filter tracking algorithms lack effective feedback mechanisms and often blindly use unreliable tracking results to update the appearance model. This causes the filter to learn a large amount of background or interference information, resulting in tracking drift or even complete failure. This phenomenon is called model contamination and is one of the key reasons for the degradation of tracker performance.
[0004] Therefore, there is an urgent need in this field for a robust target tracking method that can effectively integrate deep feature information, has anti-interference capabilities, and can perform accurate scale estimation. Summary of the Invention
[0005] Therefore, it is necessary to propose a target tracking method based on multi-layer features to address the above problems.
[0006] A target tracking method based on multi-layer features, the method comprising the following steps: Obtain the target initial position and target initial scale of the video image sequence and the first frame image in the video image sequence, and initialize the parameters of the position filter, scale filter and target feature template according to the target initial position and target initial scale of the first frame image; For the t-th frame of the video image sequence, candidate regions are determined based on the target position and target scale of the (t-1)-th frame, where t is a positive integer and its initial value is 2; The candidate region is input into a pre-trained VGG-19 network, and the depth features of the first, second, and third convolutional layers are extracted. Based on the position filter, the correlation response maps of the depth features of the first, second, and third convolutional layers are calculated respectively, which are used as the depth feature response maps of each layer. Calculate the peak-to-sidelobe ratio of the depth feature response maps in the first, second, and third convolutional layers, and adaptively fuse the weights of the first, second, and third convolutional layers based on the peak-to-sidelobe ratio to obtain the target localization response map of the t-th frame; A confidence analysis is performed on the target location response map of the t-th frame to determine the target position of the t-th frame; when the confidence level is lower than the threshold, the target position of the (t-1)-th frame is corrected, and the target position of the t-th frame is re-determined. Calculate the target scale estimate in frame t; Based on the target position in the t-th frame and the target scale estimate in the t-th frame, the parameters of the position filter, scale filter, and target feature template are updated. Let t = t + 1, then return to the step of determining the candidate region for the t-th frame of the video image sequence based on the target position and target scale of the (t-1)-th frame, until the target tracking of the video image sequence is completed.
[0007] In the above scheme, the first convolutional layer, the second convolutional layer, and the third convolutional layer are the conv3_4 layer, conv4_4 layer, and conv5_4 layer of the VGG-19 network, respectively. After extracting the depth features of the conv3_4 layer, conv4_4 layer, and conv5_4 layer, the size of the feature map output by the conv4_4 layer and the conv5_4 layer is adjusted to be consistent with the size of the feature map output by the conv3_4 layer using bilinear interpolation.
[0008] In the above scheme, calculating the peak-to-sidelobe ratio of the depth feature response maps in the first, second, and third convolutional layers specifically includes: Calculate the peak-to-sidelobe ratio of the deep feature response map using the following formula:
[0009] in, F max The maximum response value in the deep feature response map. m s1 This represents the pixel mean of the side lobe regions, excluding the peak main lobe region, in the depth feature response map. s s1 denoted as the pixel standard deviation of the sidelobe region.
[0010] In the above scheme, the adaptive fusion of the weights of the first, second, and third convolutional layers based on the peak-to-sidelobe ratio to obtain the target localization response map of frame t specifically includes: The peak-to-sidelobe ratio of the deep feature response maps of each of the three layers (conv3_4, conv4_4, and conv5_4) is calculated separately. PSR Greater than the preset threshold c PSR quantity : For each convolutional layer According to its corresponding quantity Determine the number of features used for target localization :
[0011] in, Convolutional layer Maximum quantity threshold, Convolutional layer The minimum quantity threshold; The The peak sidelobe ratio (PSR) of each feature is used as the corresponding weight, and weighted fusion is performed to obtain the final response map.
[0012] In the above scheme, the step of performing confidence analysis on the target positioning response map of the t-th frame to determine the target position in the t-th frame; and correcting the target position in the (t-1)-th frame and re-determining the target position in the t-th frame specifically includes: Calculate the average peak correlation energy (APCE) of the target localization response map as a confidence index:
[0013] Among them, F max and F min These are the maximum and minimum responses in the target localization response map, respectively. "mean" is the mean operation, and (a,b) are the pixel coordinates in the target localization response map. Set a preset energy threshold TH APCE ; If the average peak correlation energy (APCE) of the target localization response map is greater than or equal to a preset energy threshold TH APCE Then the position corresponding to the maximum response value of the target positioning response map is taken as the target position in frame t. If the average peak correlation energy (APCE) of the target localization response map is less than the preset energy threshold TH APCE If so, then an adaptive Kalman filter is used to correct the target position of the previous frame, and the corrected position is used as the target position of the t-th frame.
[0014] In the above scheme, calculating the target scale estimate in frame t specifically includes: Using the target scale of the previous frame as the central reference, construct a scale change sequence: Independent scaling filters are trained for the horizontal and vertical directions respectively, where the horizontal scaling filter is used to estimate the target width and the vertical scaling filter is used to estimate the target height; A scale reduction strategy is adopted to process the scale change sequence by dividing it into groups of three scales in sequence and retaining only the middle scale of each group to form a scale pool after dimensionality reduction. The scale pool after dimensionality reduction is initially screened using the scale filter to determine the candidate scale with the largest response value; Centered on the candidate scale, five neighboring scales are selected from the original scale change sequence for secondary discrimination. The scaling ratio corresponding to the scale with the largest response value in the secondary discrimination is taken as the target scale estimate of the t-th frame.
[0015] In the above scheme, updating the parameters of the position filter, scale filter, and target feature template based on the target position in frame t and the target scale estimate in frame t specifically includes: Based on the average peak correlation energy (APCE) and confidence threshold TH of the target localization response map in frame t. APCE The learning rate is adaptively adjusted based on the relationship between the two factors. The learning rate update formula is:
[0016] in, or l A fixed learning rate is set for the position filter. or u These are the learning rate update coefficients, used to adjust the learning rate for frame t. or t Let be the learning rate for the updated frame t; Based on the average peak correlation energy (APCE) of frame t and the threshold confidence threshold TH APCE By comparison, the learning rate update coefficients are obtained. or u：
[0017] Where m is the learning rate update constraint factor; For the target feature template and location filter parameters, a linear weighted update model is used to complete the iteration. The update formula is:
[0018] in, x t-1 and α t-1 These are the target template and filter parameters from the previous frame, respectively. x new and αnew These are the target template and filter parameters calculated for frame t, respectively. x t and α t These are the updated target template and filter parameters for frame t, respectively. or The preset adjustment coefficient; For a scaling filter, its parameters are updated in the frequency domain, and the update formula is:
[0019] in, and For the first The parameters of the scale filter after frequency domain update and These are the Fourier transforms of the desired Gaussian response and the input features, respectively. The conjugate Fourier transform of the input features. is the learning rate of the scaling filter.
[0020] This application also proposes a target tracking system based on multi-layer features, the system comprising: a data acquisition and initialization unit, a candidate region delineation unit, a deep feature extraction unit, a feature fusion unit, a target localization unit, a scale estimation unit, and an update unit; The data acquisition and initialization unit is used to acquire the target initial position and target initial scale of the video image sequence and the first frame image, and to complete the parameter initialization of the position filter, scale filter and target feature template based on the initial position and initial scale. The candidate region delineation is used to determine candidate regions for the t-th frame image in the video image sequence based on the target position and target scale of the (t-1)-th frame, where t is a positive integer and its initial value is 2; The deep feature extraction unit has a built-in pre-trained VGG-19 network, which is used to input the candidate region into the VGG-19 network, extract the depth features of the first convolutional layer, the second convolutional layer and the third convolutional layer in the network, and calculate the correlation response maps of the depth features of the first convolutional layer, the second convolutional layer and the third convolutional layer based on the position filter, respectively, as the depth feature response maps of each layer. The feature fusion unit is used to calculate the peak-to-sidelobe ratio of the depth feature response map of the three convolutional layers, and then adaptively fuse the feature weights of the three convolutional layers based on the peak-to-sidelobe ratio to output the target localization response map of the t-th frame. The target positioning unit is used to perform confidence analysis on the target positioning response map and determine the target position in frame t based on the analysis results; when the confidence is lower than a preset threshold, the target position in frame t-1 is corrected and the target position in frame t is re-output. The scale estimation unit is used to calculate the target scale estimate in frame t. The updating unit is used to update the parameters of the position filter, the scale filter, and the target feature template based on the target position in the t-th frame and the target scale estimate in the t-th frame. Let t = t + 1, then return to the step of determining the candidate region for the t-th frame of the video image sequence based on the target position and target scale of the (t-1)-th frame, until the target tracking of the video image sequence is completed.
[0021] This application also proposes a readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the following steps: Obtain the target initial position and target initial scale of the video image sequence and the first frame image in the video image sequence, and initialize the parameters of the position filter, scale filter and target feature template according to the target initial position and target initial scale of the first frame image; For the t-th frame of the video image sequence, candidate regions are determined based on the target position and target scale of the (t-1)-th frame, where t is a positive integer and its initial value is 2; The candidate region is input into a pre-trained VGG-19 network, and the depth features of the first, second, and third convolutional layers are extracted. Based on the position filter, the correlation response maps of the depth features of the first, second, and third convolutional layers are calculated respectively, which are used as the depth feature response maps of each layer. Calculate the peak-to-sidelobe ratio of the depth feature response maps in the first, second, and third convolutional layers, and adaptively fuse the weights of the first, second, and third convolutional layers based on the peak-to-sidelobe ratio to obtain the target localization response map of the t-th frame; A confidence analysis is performed on the target location response map of the t-th frame to determine the target position of the t-th frame; when the confidence level is lower than the threshold, the target position of the (t-1)-th frame is corrected, and the target position of the t-th frame is re-determined. Calculate the target scale estimate in frame t; Based on the target position in the t-th frame and the target scale estimate in the t-th frame, the parameters of the position filter, scale filter, and target feature template are updated. Let t = t + 1, then return to the step of determining the candidate region for the t-th frame of the video image sequence based on the target position and target scale of the (t-1)-th frame, until the target tracking of the video image sequence is completed.
[0022] This application also proposes a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor causes the processor to perform the following steps: Obtain the target initial position and target initial scale of the video image sequence and the first frame image in the video image sequence, and initialize the parameters of the position filter, scale filter and target feature template according to the target initial position and target initial scale of the first frame image; For the t-th frame of the video image sequence, candidate regions are determined based on the target position and target scale of the (t-1)-th frame, where t is a positive integer and its initial value is 2; The candidate region is input into a pre-trained VGG-19 network, and the depth features of the first, second, and third convolutional layers are extracted. Based on the position filter, the correlation response maps of the depth features of the first, second, and third convolutional layers are calculated respectively, which are used as the depth feature response maps of each layer. Calculate the peak-to-sidelobe ratio of the depth feature response maps in the first, second, and third convolutional layers, and adaptively fuse the weights of the first, second, and third convolutional layers based on the peak-to-sidelobe ratio to obtain the target localization response map of the t-th frame; A confidence analysis is performed on the target location response map of the t-th frame to determine the target position of the t-th frame; when the confidence level is lower than the threshold, the target position of the (t-1)-th frame is corrected, and the target position of the t-th frame is re-determined. Calculate the target scale estimate in frame t; Based on the target position in the t-th frame and the target scale estimate in the t-th frame, the parameters of the position filter, scale filter, and target feature template are updated. Let t = t + 1, then return to the step of determining the candidate region for the t-th frame of the video image sequence based on the target position and target scale of the (t-1)-th frame, until the target tracking of the video image sequence is completed.
[0023] The embodiments of this invention offer the following advantages: By utilizing the VGG-19 deep network to extract depth features at different levels, this invention combines the low-level detailed texture information of the target with high-level semantic information, overcoming the limitations of hand-crafted features. Simultaneously, it introduces Peak Sidelobe Ratio (PSR) as an indicator of feature channel quality, and accordingly performs adaptive channel filtering and weighted fusion. This method automatically assigns greater weight to higher-quality, more reliable feature channels, thereby significantly improving the discriminative power of feature representation. Furthermore, confidence analysis is performed on the target localization response map of frame t. When the confidence level is low, a target relocalization mechanism is triggered, avoiding the direct use of unreliable localization results and greatly improving the accuracy of the target tracking algorithm. Attached Figure Description
[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0025] in: Figure 1 This is a schematic diagram of a target tracking method based on multi-layer features in one embodiment; Figure 2(a) shows the Bird2 sequence image; Figure 2(b) shows the Couple sequence image; Figure 2(c) shows the Dog1 sequence image; Figure 2(d) shows the DragonBaby sequence image; Figure 2(e) shows the FootBall sequence image; Figure 2(f) shows the Skiing sequence image; Figure 2(g) shows the fish3 sequence image; Figure 2(h) shows the pedestrian2 sequence image. Detailed Implementation
[0026] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0027] In the following description, numerous specific details are set forth in order to provide a more thorough understanding of the invention; however, it will be apparent to those skilled in the art that the invention may be practiced without one or more of these details; in other instances, certain technical features well-known in the art have not been described in order to avoid confusion with the invention. It should be understood that the invention can be practiced in different forms and should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided to make the disclosure thorough and complete and to fully convey the scope of the invention to those skilled in the art.
[0028] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. When used herein, the singular forms “a,” “an,” and “the” are also intended to include the plural forms, unless the context clearly indicates otherwise. The terms “comprising” and / or “including,” when used in this specification, identify the presence of said features, integers, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups. When used herein, the term “and / or” includes any and all combinations of the associated listed items.
[0029] To fully understand the present invention, a detailed structure will be presented in the following description in order to illustrate the technical solution proposed by the present invention; optional embodiments of the present invention are described in detail below, however, in addition to these detailed descriptions, the present invention may have other embodiments.
[0030] like Figure 1 As shown, in one embodiment, a target tracking method based on multi-layer features is provided. This multi-layer feature-based target tracking method includes steps S101 to S107, which are detailed below: S101. Obtain the initial target position and initial target scale of the video image sequence and the first frame image in the video image sequence, and initialize the parameters of the position filter, scale filter and target feature template according to the initial target position and initial target scale of the first frame image. By acquiring target information from the first frame—position and size—an initial target feature template and filters for predicting position and scale can be established. Accurate initialization is crucial to ensuring the entire tracking process is on track from the very beginning.
[0031] S102. For the t-th frame image in the video image sequence, determine the candidate region based on the target position and target scale of the (t-1)-th frame, where t is a positive integer and the initial value is 2. This step is based on the assumption that the target's movement between consecutive frames is usually smooth and continuous. A reasonable candidate region is defined centered on the target's position in the previous frame. This allows subsequent feature extraction and computation to be concentrated within a small region, greatly improving the algorithm's real-time performance.
[0032] S103. Input the candidate region into the pre-trained VGG-19 network, extract the depth features of the first convolutional layer, the second convolutional layer and the third convolutional layer, and calculate the correlation response maps of the depth features of the first convolutional layer, the second convolutional layer and the third convolutional layer based on the position filter, as the depth feature response maps of each layer. The VGG-19 network's powerful feature extraction capabilities can simultaneously capture detailed textures from shallow layers and high-level semantic information from deeper layers. By performing correlation operations between positional filters and features from each layer, a response map—the vote of each feature layer on the target location—can be obtained. This provides a multi-dimensional information foundation for subsequent precise localization.
[0033] In some embodiments, the first convolutional layer, the second convolutional layer, and the third convolutional layer are the conv3_4 layer, conv4_4 layer, and conv5_4 layer of the VGG-19 network, respectively. After extracting the depth features of the conv3_4 layer, conv4_4 layer, and conv5_4 layer, bilinear interpolation is used to adjust the size of the feature map output by the conv4_4 layer to be consistent with the size of the feature map output by the conv3_4 layer.
[0034] Specifically, comparing the feature maps output by three different convolutional layers (conv3_4, conv4_4, and conv5_4), generally, the feature map obtained through the conv3_4 convolutional layer preserves edge features better, while the feature map obtained through the conv5_4 convolutional layer, although the shape of the target is barely discernible, has a significantly higher contrast between the target and the background. If the size of the output feature maps from the three convolutional layers decreases in that order, the feature map output by the conv5_4 layer is only 14×14 pixels, which is insufficient to distinguish the target's location. Therefore, bilinear interpolation is used to adjust the feature maps output by the conv4_4 and conv5_4 layers to the same size as the feature map output by the conv3_4 layer. The method is as follows:
[0035] in, x i Let i be the feature vector at the i-th position after bilinear interpolation. Let j be the feature vector at the j-th position of the original feature map. m ij These are the interpolation weights.
[0036] S104. Calculate the peak-to-sidelobe ratio of the depth feature response maps in the first, second, and third convolutional layers. Based on the peak-to-sidelobe ratio, adaptively fuse the weights of the first, second, and third convolutional layers to obtain the target localization response map of the t-th frame. Specifically, the Peak-to-Side-Lobe Ratio (PSR) is an indicator of response map quality—a higher PSR indicates a more prominent response peak and more reliable localization results. By adaptively fusing PSR as weights, the algorithm can automatically amplify the role of high-quality features while suppressing distorted or poor-quality features, thereby generating a more accurate and robust final localization response map than any single feature layer.
[0037] In some embodiments, calculating the peak-to-sidelobe ratio of the deep feature response maps in the first, second, and third convolutional layers specifically includes: Calculate the peak-to-sidelobe ratio of the deep feature response map using the following formula:
[0038] in, F max The maximum response value in the deep feature response map. m s1 This represents the pixel mean of the side lobe regions, excluding the peak main lobe region, in the depth feature response map. s s1 denoted as the pixel standard deviation of the sidelobe region.
[0039] Preferably, if the number of channels with a PSR value greater than 20 in a certain convolutional layer is greater than or equal to the maximum channel count threshold, then the first... n imax One channel feature is used for target localization; if the number of channels with a PSR value greater than 20 is less than the maximum channel number threshold, but greater than or equal to the minimum channel number threshold, then all channels with a PSR value greater than 20 are selected; if the number of channels with a PSR value greater than 20 is less than the minimum channel number threshold, then the first few channels are selected. n imin Each channel feature.
[0040] in, β Channel number threshold ratio: The number of channels specifically refers to the number of effective feature channels used for final fusion after screening.
[0041] In some embodiments, the weights of the first, second, and third convolutional layers are adaptively fused based on the peak-sidelobe ratio to obtain the target localization response map of frame t, specifically including: Statistically analyze the peak-to-sidelobe ratio of the deep feature response maps of each layer (conv3_4, conv4_4, and conv5_4). PSR Greater than the preset threshold c PSR quantity : For each convolutional layer According to its corresponding quantity Determine the number of features used for target localization :
[0042] in, Convolutional layer Maximum quantity threshold, Convolutional layer The minimum quantity threshold; Will The peak sidelobe ratio (PSR) of each feature is used as the corresponding weight, and weighted fusion is performed to obtain the final response map.
[0043] Preferably, the PSR values of all selected channels are used as their respective fusion weights and normalized to obtain the fusion weight of the i-th channel as follows:
[0044] in, Let i be the fusion weight of the i-th channel. Let be the peak-to-sidelobe ratio of the i-th channel. This is the PSR value of the k-th channel.
[0045] The final response after fusion is:
[0046] in, To locate the final target response map, Let i be the fusion weight of the i-th channel. This is the characteristic response map of the i-th channel.
[0047] Specifically, the final target localization response map is a two-dimensional matrix, where the value of each pixel represents the probability score of being the target center at that location. The location of the maximum value in the map is determined as the target location in the current frame. This is achieved by integrating multi-channel information and weighting the integration based on reliability. The response maps of high-quality channels are amplified, while the influence of channels that may contain noise or have weak discriminative power is suppressed. The final response map is expected to have sharper peaks, more accurate locations, and stronger anti-interference capabilities, thereby directly improving the accuracy and robustness of target localization.
[0048] S105. Perform confidence analysis on the target location response map of frame t to determine the target position of frame t; when the confidence is lower than the threshold, correct the target position of frame t-1 and redetermine the target position of frame t. This step assesses the reliability of the current localization result and initiates an error correction mechanism when it becomes unreliable to prevent tracking failure. When the target is occluded, undergoes severe deformation, or is interfered with by similar objects, the quality of the fused response map decreases, resulting in low confidence. In such cases, this scheme uses a correction mechanism to predict a more reasonable target location. This effectively avoids tracking drift and significantly improves the algorithm's robustness in complex scenarios.
[0049] In some embodiments, a confidence analysis is performed on the target localization response map of frame t to determine the target position in frame t; when the confidence level is lower than a threshold, the target position in frame t-1 is corrected, and the target position in frame t is re-determined, specifically including: Calculate the average peak correlation energy (APCE) of the target localization response map as a confidence index:
[0050] Among them, F max and F min These are the maximum and minimum responses in the target localization response map, respectively. "mean" is the mean operation, and (a, b) are the pixel coordinates in the target localization response map. Set a preset energy threshold TH APCE ; If the average peak correlation energy (APCE) of the target localization response map is greater than or equal to the preset energy threshold TH APCE Then the position corresponding to the maximum response value of the target localization response map is taken as the target position in frame t. If the average peak correlation energy (APCE) of the target localization response map is less than the preset energy threshold TH APCE If so, then an adaptive Kalman filter is used to correct the target position of the previous frame, and the corrected position is used as the target position of the t-th frame.
[0051] Among them, if the average peak correlation energy (APCE) value of the target localization response map is less than the preset energy threshold TH APCE If the target is not properly tracked, it indicates that the target may be affected by deformation, occlusion, or interference from similar objects, and is in an abnormal tracking state. In this case, adaptive Kalman filtering is required to perform anti-interference localization of the target and reacquire the target position in the current frame.
[0052] Preferably, the APCE value can reflect the distribution of the relevant response well. A threshold THAPCE is set based on the mean of historical frame APCE values, which can be expressed as:
[0053] Where t is the frame number of the current frame. i APCE As a threshold parameter, after experimentation, i APCEA value of 0.55 yields better results. PACE i Let be the APCE value of the i-th frame.
[0054] S106. Calculate the target scale estimate in frame t; In real-world scenarios, the distance between the target and the camera varies, causing its scale in the image to change accordingly. Through independent scale estimation, the algorithm can adjust the size of the target bounding box in real time, ensuring that it always tightly surrounds the target. This is crucial for maintaining tracking accuracy and providing accurate candidate regions for the next frame.
[0055] In some embodiments, calculating the target scale estimate for frame t specifically includes: Using the target scale of the previous frame as the central reference, construct a scale change sequence: Independent scaling filters are trained for the horizontal and vertical directions respectively, where the horizontal scaling filter is used to estimate the target width and the vertical scaling filter is used to estimate the target height; A scale reduction strategy is adopted to process the data. The scale change sequence is divided into groups of three scales in sequence, and only the middle scale of each group is retained to form a scale pool after dimensionality reduction. The scale pool after dimensionality reduction is initially screened by a scale filter to determine the candidate scale with the largest response value; Centered on the candidate scale, five neighboring scales are selected from the original scale change sequence for secondary discrimination. The scaling ratio corresponding to the scale with the largest response value in the secondary discrimination is taken as the target scale estimate of the t-th frame.
[0056] In some embodiments, to perform accurate scale estimation of the target, a scale filter based on a scale pyramid is first introduced, and its first... t When performing frame detection, the obtained target position is used as the center, and the frame is... t The target size in frame -1 is used as the center scale, and then scaled proportionally. S Image patches of several scales are used as the images to be detected, and the image size of each scale layer is:
[0057] in, a As a scaling factor, it is generally taken as... a =1.02.
[0058] The obtained S Each image was resized using bilinear interpolation. n The image size when =0 is calculated, and the filter responses are obtained respectively. The scale corresponding to the maximum response is used as the scale of the current frame; the maximum response can be expressed as:
[0059] Since the scaling filter scales the target based on the aspect ratio of the first frame, its aspect ratio will change when the target rotates or deforms, leading to some error in the scaling process. To address this, this algorithm trains the aforementioned scaling filter separately for the horizontal and vertical directions of the image, applying bidirectional filters to calculate the target's width and height respectively.
[0060] For a filter in a certain direction, the basic scale layer number is generally taken. S =33, then its scale pool is:
[0061] For this scale pool, if we group three scales sequentially and retain only the middle scale of each group, then the scale pool will become:
[0062] After scale filtering in a certain direction, the original five scale layers are used as the center for secondary discrimination, centered on the scale layer with the largest response value. For example, if the scale layer with the largest response value in a certain frame is -9, then the scale pool for secondary discrimination should be:
[0063] Since the response values of the -12 and -6 scale layers were determined to be lower than those of the 9 scale layer in the initial discrimination, this method can achieve scale dimensionality reduction while ensuring the accuracy of scale estimation. This reduces the number of scale layers processed in the original single direction from 33 to 16, effectively improving computational efficiency.
[0064] In addition, if the scale layer with the largest response value in a certain frame is -15 or 15, then the scale pool for secondary discrimination processing only needs to be [-16, -15, -14, -13] or [13, 14, 15, 16]. In this case, the number of scale layers to be processed for a single direction is 15.
[0065] Preferably, the method for judging the target scale change and determining the order of horizontal and vertical scale calculation is as follows: In applying a bidirectional filter, the degree of scale change of the target in the previous frame is calculated based on the scale of the previous two frames:
[0066] in, w and h These are the width and height of the target, respectively; If the degree of variation in the width and height of the image meets the following conditions: ; This indicates that in the first tIn frame -1, the image width changes significantly, so horizontal scaling is performed first, while the vertical scaling value is fixed at the height of the target from the previous frame. h t-1 Remain unchanged, and proceed in the same manner as before. S Filtering is performed on each scale layer, and the scale layer with the largest response is selected. w Target width of the current frame w t Then, vertical scaling is performed, while the horizontal scaling value is fixed at the width of the target in the current frame. w t Without changing the settings, perform the same filtering process to obtain the target height of the current frame. h t .
[0067] Similarly, if in the first t If the image height changes significantly within frame -1, then vertical scaling must be performed first, followed by horizontal scaling.
[0068] S107. Update the parameters of the position filter, scale filter, and target feature template based on the target position in frame t and the target scale estimate in frame t. By using the target's location and scale confirmed in the current frame, the algorithm can update its internal filters and templates. This prevents the model from becoming outdated and allows for continuous and accurate target tracking. Furthermore, combining confidence analysis enables adaptive updates; for example, the learning rate can be reduced when confidence is low, preventing the model from being contaminated by erroneous information when subjected to interference.
[0069] Preferably, a scaling filter based on a scaling pyramid is introduced. First, the minimum error function is constructed as follows:
[0070] in, f As a feature, g It is a two-dimensional Gaussian function. h For filters, d For feature dimensions.
[0071] By introducing the above equation into the frequency domain and minimizing it, we can obtain the filter as follows:
[0072] in, and They are respectively g and f The conjugate Fourier transform of .
[0073] In some embodiments, the target state equation is updated according to the following steps: the target's motion state vector is defined as follows.x =[ x , y , dx , day ] T ,in( x , y () is the center location of the target. dx and day Based on the target's velocities in the horizontal and vertical directions, a process model of the target can be established, which can be represented as:
[0074] in, A k and B k These are the state transition matrix and the control matrix, respectively. x k express k The true value of the target at any given time, representing the process model. u k-1 for k The control input at time -1 w k The noise is the process noise, with a mean of 0 and a covariance matrix of... Q k Gaussian noise distribution.
[0075] The observation model for the target can be represented as:
[0076] in, z k For the observed values, H k To observe the transition matrix, v k The observation noise has a mean of 0 and a covariance matrix of... R k The Gaussian noise distribution is such that the two noises are independent of each other.
[0077] Predict the target state information in the current frame based on the best predicted target state information from the previous frame:
[0078] in, The predicted value of the current process model is called the prior estimate; The update method for the covariance of the process model is as follows:
[0079] in, The error covariance matrix is predicted from the error covariance matrix and the process noise covariance matrix of the previous process. Q k Let be the covariance matrix.
[0080] The predicted values are updated and optimized using the observed values, resulting in the observation update equation:
[0081] in, x k This indicates that after the update, in k The optimal predicted value at time 10:00. K k Represents the Kalman gain matrix. P k The updated error covariance matrix, I It is the identity matrix; Based on the noise covariance of the prediction error dynamic correction process Q and observation noise covariance matrix R The specific method is as follows
[0082] in, Q i and R i These are the process noise covariance matrices. Q and observation noise covariance matrix R The elements in P i and P pre These are the state covariances of the current state and the previous state, respectively. z i and These are the actual and predicted observations, respectively.
[0083] In some embodiments, the parameters of the position filter, scale filter, and target feature template are updated based on the target position in frame t and the target scale estimate in frame t, specifically including: Based on the average peak correlation energy (APCE) and confidence threshold TH of the target localization response map in frame t. APCE The learning rate is adaptively adjusted based on the relationship between the two factors. The learning rate update formula is:
[0084] in, or l A fixed learning rate is set for the position filter. or u These are the learning rate update coefficients, used to adjust the learning rate for frame t. or t Let be the learning rate for the updated frame t; Based on the average peak correlation energy (APCE) of frame t and the threshold confidence threshold TH APCE By comparison, the learning rate update coefficients are obtained. or u：
[0085] Where m is the learning rate update constraint factor; For the target feature template and location filter parameters, a linear weighted update model is used to complete the iteration. The update formula is:
[0086] in, x t-1 and α t-1 These are the target template and filter parameters from the previous frame, respectively. x new and α new These are the target template and filter parameters calculated for frame t, respectively. x t and α t These are the updated target template and filter parameters for frame t, respectively. or The preset adjustment coefficient; For a scaling filter, its parameters are updated in the frequency domain, and the update formula is:
[0087] in, and For the first The parameters of the scale filter after frequency domain update and These are the Fourier transforms of the desired Gaussian response and the input features, respectively. The conjugate Fourier transform of the input features. is the learning rate of the scaling filter.
[0088] Specifically, when PACE Value greater than threshold TH APCE When the target is in a certain state, it is considered normal tracking, and the filter and target feature template are updated normally; when... PACE Value less than threshold TH APCE However, it is greater than the threshold. TH APCEWhen the learning rate is 0.8 times, the target may be affected by deformation, occlusion, and interference, causing the filter to learn irrelevant features. Therefore, it is appropriate to reduce the learning rate. PACE Value less than threshold TH APCE When the learning rate is 0.8 times, it indicates that the target is affected by severe deformation, occlusion and interference. The filter and target feature template are easily contaminated by irrelevant features such as occluded areas. Therefore, the learning rate is set to 0, that is, the result of this frame does not participate in the update of the filter and target feature template.
[0089] S108. Let t = t + 1, return to the step of determining the candidate region for the t-th frame of the video image sequence based on the target position and target scale of the (t-1)-th frame, until the target tracking of the video image sequence is completed.
[0090] In summary, by continuously repeating this cycle, it is possible to achieve continuous and dynamic tracking of targets in a video sequence.
[0091] In some embodiments, the anomaly detection effect of the present invention will be further explained below with reference to FIG2.
[0092] 1. Simulation conditions: To verify the effectiveness of the proposed feature adaptive fusion tracking algorithm, this section uses the OTB-100 and VOT datasets as benchmark datasets and compares the KCF, SAMF, FDSST, MUSTer, and CSR-DCF algorithms to evaluate the performance of the proposed feature adaptive fusion tracking algorithm.
[0093] 2. Simulation Results and Analysis: Figure 2 shows the results of this invention. The algorithm in this chapter and the comparison algorithm were tested on the same dataset sequence to compare the results and illustrate the advancement of this invention.
[0094] Figure 2(a) shows the Bird2 sequence images, with the target being a white crane in the lower left corner of the image. At frame 50, the target suddenly changed from moving right to left, causing a certain degree of offset in the tracking results of the CSR-DCF, KCF, and FDSST algorithms. At frame 77, the CSR-DCF and KCF algorithms still exhibited tracking offset issues, while the FDSST and SAMF algorithms incorrectly tracked the target, mistaking it for a fledgling. At frame 99, the image shifted significantly, causing the target to move considerably and become blurred. The MUSTer algorithm also lost track of the target, while only the algorithm of this invention maintained correct tracking throughout the entire sequence.
[0095] Figure 2(b) shows a couple sequence of images, with the pedestrians in the image being the tracking target. It's important to note that the target in this sequence is the entire pair of pedestrians, not just one of them, as can be seen from frame 1. At frame 35, the KCF algorithm experiences tracking drift, leading to tracking failure. At frame 47, due to rapid camera movement, the MUSTer, SAMF, and FDSST algorithms all experience tracking drift. However, the MUSTer algorithm, with its long-term memory, recovers the tracking target in subsequent frames, while the SAMF and FDSST algorithms fail. At frame 131, the MUSTer algorithm's tracking target changes to the pedestrian on the left, while the CSR-DCF algorithm's target changes to the pedestrian on the right. Only the algorithm of this invention maintains correct tracking.
[0096] Figure 2(c) shows the Dog1 sequence images, with the head of a toy dog in the image being the tracked target. At frame 183, the CSR-DCF algorithm exhibits tracking scale issues; the lower boundary of the tracking box shifts downwards relative to the standard case, and the right boundary shifts to the right relative to the standard case. At frame 820, due to the target rapidly approaching the camera, the KCF, MUSTer, and SAMF algorithms also experience tracking scale problems, with the tracked area being smaller than the target. Among these, the KCF algorithm, lacking a scale adjustment module, shows the largest discrepancy between its scale and the actual target. Frame 929 provides a more intuitive view of the tracking scale errors of the CSR-DCF, KCF, MUSTer, and SAMF algorithms; only the algorithm of this invention and the FDSST algorithm show good tracking performance.
[0097] Figure 2(d) shows the DragonBaby image sequence, with the target being the head of the child on the left side of the image. In frame 28, the target turns around, causing facial features to be lost, resulting in the failure of the KCF and FDSST algorithms. In frame 47, due to the target's rapid movement, the other algorithms experienced significant deviations, while only the algorithm presented in this paper maintained good tracking performance. In frame 81, due to a sharp change in the image's field of view, the other algorithms lost track of the target, while only the algorithm of this invention continued to track correctly.
[0098] Figure 2(e) shows the FootBall sequence images, with the tracking target being the head of player number 24 in the image. At frame 200, the CSR-DCF, KCF, MUSTer, and FDSST algorithms all exhibited tracking offsets. At frame 289, the head of another player, number 37, obscured the tracking target, and the two players shared similar features, causing the CSR-DCF, KCF, MUSTer, and SAMF algorithms to track the wrong target. By the end of frame 362, the FDSST algorithm's tracking result had shifted slightly to the left, while only the algorithm of this invention maintained stable and accurate tracking.
[0099] Figure 2(f) shows a Skiing sequence image, with the skier in the image being the tracking target. In frame 9, the target begins to move in the air, causing the KCF, MUSTer, fDSST, and SAMF algorithms to lose track of the target; only the algorithm described in this paper and the CSR-DCF algorithm maintain tracking. In frame 14, the target continues to move rapidly in the air, at which point the CSR-DCF algorithm also loses track. In frame 58, the target lands, and the algorithm of this invention still maintains stable and accurate tracking.
[0100] Figure 2(g) shows the fish3 sequence image, with the yellow fish in the image being tracked. At frame 42, the MUSTer algorithm lost track of the target due to its rapid movement. At frame 311, the FDSST algorithm gradually lost track of the target, and the CSR-DCF algorithm experienced a significant shift in its tracking results. At frame 519, due to the target's rotation, the KCF and SAMF algorithms failed to adapt to the target's scale changes; only the algorithm of this invention maintained relatively stable and accurate tracking.
[0101] Figure 2(h) shows a pedestrian2 image sequence, with the target being tracked. At frame 177, due to the small size and low resolution of the target, the KCF and SAMF algorithms lost track of it. At frame 495, the CSR-DCF and MUSTer algorithms showed significant scale deviations in their results. By the end of frame 713, only the algorithm presented in this paper and the fDSST algorithm completed tracking, with the fDSST algorithm showing a slight downward-right offset, while the algorithm presented in this paper achieved more accurate tracking results.
[0102] This application also proposes a target tracking system based on multi-layer features, which includes: a data acquisition and initialization unit, a candidate region delineation unit, a deep feature extraction unit, a feature fusion unit, a target localization unit, a scale estimation unit, and an update unit; The data acquisition and initialization unit is used to acquire the target initial position and target initial scale of the video image sequence and the first frame image, and to complete the parameter initialization of the position filter, scale filter and target feature template based on the initial position and initial scale. Candidate region delineation is used to determine candidate regions for the t-th frame of a video image sequence based on the target position and target scale of the (t-1)-th frame, where t is a positive integer and its initial value is 2; The deep feature extraction unit has a built-in pre-trained VGG-19 network. It is used to input candidate regions into the VGG-19 network and extract the deep features of the first, second, and third convolutional layers in the network. Based on the position filter, it calculates the correlation response maps of the deep features of the first, second, and third convolutional layers respectively, which are used as the deep feature response maps of each layer. The feature fusion unit is used to calculate the peak-to-sidelobe ratio of the depth feature response map of the three convolutional layers, and then adaptively fuse the feature weights of the three convolutional layers based on the peak-to-sidelobe ratio to output the target localization response map of the t-th frame. The target localization unit is used to perform confidence analysis on the target localization response map and determine the target position in frame t based on the analysis results. When the confidence is lower than a preset threshold, the target position in frame t-1 is corrected and the target position in frame t is re-output. The scale estimation unit is used to calculate the target scale estimate in frame t. The update unit is used to update the parameters of the position filter, scale filter, and target feature template based on the target position and target scale estimate in frame t. Let t = t + 1, then return to the step of determining the candidate region based on the target position and target scale of the target in the t-1 frame of the video image sequence, until the target tracking of the video image sequence is completed.
[0103] This application also proposes a readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the following steps: Obtain the initial target position and initial target scale of the video image sequence and the first frame image in the video image sequence, and initialize the parameters of the position filter, scale filter and target feature template based on the initial target position and initial target scale of the first frame image; For the t-th frame of a video image sequence, candidate regions are determined based on the target location and target scale of the (t-1)-th frame, where t is a positive integer and its initial value is 2; The candidate region is input into the pre-trained VGG-19 network, and the depth features of the first, second and third convolutional layers are extracted. Based on the position filter, the correlation response maps of the depth features of the first, second and third convolutional layers are calculated respectively, which are used as the depth feature response maps of each layer. Calculate the peak-to-sidelobe ratio of the depth feature response maps in the first, second, and third convolutional layers, and adaptively fuse the weights of the first, second, and third convolutional layers based on the peak-to-sidelobe ratio to obtain the target localization response map of the t-th frame; Confidence analysis is performed on the target localization response map of frame t to determine the target position of frame t; when the confidence is lower than the threshold, the target position of frame t-1 is corrected and the target position of frame t is re-determined. Calculate the target scale estimate in frame t; Based on the target position and target scale estimate in frame t, the parameters of the position filter, scale filter, and target feature template are updated. Let t = t + 1, then return to the step of determining the candidate region based on the target position and target scale of the target in the t-1 frame of the video image sequence, until the target tracking of the video image sequence is completed.
[0104] This application also proposes a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the following steps: Obtain the initial target position and initial target scale of the video image sequence and the first frame image in the video image sequence, and initialize the parameters of the position filter, scale filter and target feature template based on the initial target position and initial target scale of the first frame image; For the t-th frame of a video image sequence, candidate regions are determined based on the target location and target scale of the (t-1)-th frame, where t is a positive integer and its initial value is 2; The candidate region is input into the pre-trained VGG-19 network, and the depth features of the first, second and third convolutional layers are extracted. Based on the position filter, the correlation response maps of the depth features of the first, second and third convolutional layers are calculated respectively, which are used as the depth feature response maps of each layer. Calculate the peak-to-sidelobe ratio of the depth feature response maps in the first, second, and third convolutional layers, and adaptively fuse the weights of the first, second, and third convolutional layers based on the peak-to-sidelobe ratio to obtain the target localization response map of the t-th frame; Confidence analysis is performed on the target localization response map of frame t to determine the target position of frame t; when the confidence is lower than the threshold, the target position of frame t-1 is corrected and the target position of frame t is re-determined. Calculate the target scale estimate in frame t; Based on the target position and target scale estimate in frame t, the parameters of the position filter, scale filter, and target feature template are updated. Let t = t + 1, then return to the step of determining the candidate region based on the target position and target scale of the target in the t-1 frame of the video image sequence, until the target tracking of the video image sequence is completed.
[0105] Those skilled in the art will understand that implementing all or part of the processes in the above embodiments can be accomplished by instructing related hardware through a computer program. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.
[0106] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0107] The embodiments described above are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of this application's patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. The embodiments disclosed above are merely preferred embodiments of the present invention and should not be construed as limiting the scope of the present invention. Therefore, equivalent variations made according to the claims of this invention are still within the scope of this invention.
Claims
1. A target tracking method based on multi-layer features, characterized in that, The method includes: Obtain the target initial position and target initial scale of the video image sequence and the first frame image in the video image sequence, and initialize the parameters of the position filter, scale filter and target feature template according to the target initial position and target initial scale of the first frame image; For the t-th frame of the video image sequence, candidate regions are determined based on the target position and target scale of the (t-1)-th frame, where t is a positive integer and its initial value is 2; The candidate region is input into a pre-trained VGG-19 network, and the depth features of the first, second, and third convolutional layers are extracted. Based on the position filter, the correlation response maps of the depth features of the first, second, and third convolutional layers are calculated respectively, which are used as the depth feature response maps of each layer. Calculate the peak-to-sidelobe ratio of the depth feature response maps in the first, second, and third convolutional layers, and adaptively fuse the weights of the first, second, and third convolutional layers based on the peak-to-sidelobe ratio to obtain the target localization response map of the t-th frame; A confidence analysis is performed on the target location response map of the t-th frame to determine the target position of the t-th frame; when the confidence level is lower than the threshold, the target position of the (t-1)-th frame is corrected, and the target position of the t-th frame is re-determined. Calculate the target scale estimate in frame t; Based on the target position in the t-th frame and the target scale estimate in the t-th frame, the parameters of the position filter, scale filter, and target feature template are updated. Let t = t + 1, then return to the step of determining the candidate region for the t-th frame of the video image sequence based on the target position and target scale of the (t-1)-th frame, until the target tracking of the video image sequence is completed.
2. The target tracking method based on multi-layer features according to claim 1, characterized in that, The first, second, and third convolutional layers are the conv3_4, conv4_4, and conv5_4 layers of the VGG-19 network, respectively. After extracting the depth features of the conv3_4, conv4_4, and conv5_4 layers, bilinear interpolation is used to adjust the size of the feature maps output by the conv4_4 and conv5_4 layers to be consistent with the size of the feature map output by the conv3_4 layer.
3. The target tracking method based on multi-layer features according to claim 2, characterized in that, The calculation of the peak-to-sidelobe ratio of the depth feature response maps in the first, second, and third convolutional layers specifically includes: Calculate the peak-to-sidelobe ratio of the deep feature response map using the following formula: in, F max The maximum response value in the deep feature response map. μ s1 This represents the pixel mean of the side lobe regions, excluding the peak main lobe region, in the depth feature response map. σ s1 denoted as the pixel standard deviation of the sidelobe region.
4. The target tracking method based on multi-layer features according to claim 3, characterized in that, The adaptive fusion of the weights of the first, second, and third convolutional layers based on the peak-sidelobe ratio to obtain the target localization response map for frame t specifically includes: The peak-to-sidelobe ratio of the deep feature response maps of each of the three layers (conv3_4, conv4_4, and conv5_4) is calculated separately. PSR Greater than the preset threshold c PSR quantity : For each convolutional layer According to its corresponding quantity Determine the number of features used for target localization : in, Convolutional layer Maximum quantity threshold, Convolutional layer The minimum quantity threshold; The The peak sidelobe ratio (PSR) of each feature is used as the corresponding weight, and weighted fusion is performed to obtain the final response map.
5. The target tracking method based on multi-layer features according to claim 4, characterized in that, The confidence analysis is performed on the target localization response map of the t-th frame to determine the target location in the t-th frame; When the confidence level is lower than the threshold, the target position in frame t-1 is corrected, and the target position in frame t is redefined, specifically including: Calculate the average peak correlation energy (APCE) of the target localization response map as a confidence index: Among them, F max and F min These are the maximum and minimum responses in the target localization response map, respectively. "mean" is the mean operation, and (a,b) are the pixel coordinates in the target localization response map. Set a preset energy threshold TH APCE ; If the average peak correlation energy (APCE) of the target localization response map is greater than or equal to a preset energy threshold TH APCE Then the position corresponding to the maximum response value of the target positioning response map is taken as the target position in frame t. If the average peak correlation energy (APCE) of the target localization response map is less than the preset energy threshold TH APCE If so, then an adaptive Kalman filter is used to correct the target position of the previous frame, and the corrected position is used as the target position of the t-th frame.
6. The target tracking method based on multi-layer features according to claim 5, characterized in that, The calculation of the target scale estimate in frame t specifically includes: Using the target scale of the previous frame as the central reference, construct a scale change sequence: Independent scaling filters are trained for the horizontal and vertical directions respectively, where the horizontal scaling filter is used to estimate the target width and the vertical scaling filter is used to estimate the target height; A scale reduction strategy is adopted to process the scale change sequence by dividing it into groups of three scales in sequence and retaining only the middle scale of each group to form a scale pool after dimensionality reduction. The scale pool after dimensionality reduction is initially screened using the scale filter to determine the candidate scale with the largest response value; Centered on the candidate scale, five neighboring scales are selected from the original scale change sequence for secondary discrimination. The scaling ratio corresponding to the scale with the largest response value in the secondary discrimination is taken as the target scale estimate of the t-th frame.
7. The target tracking method based on multi-layer features according to claim 6, characterized in that, The step of updating the parameters of the position filter, scale filter, and target feature template based on the target position in frame t and the target scale estimate in frame t specifically includes: Based on the average peak correlation energy (APCE) and confidence threshold TH of the target localization response map in frame t. APCE The learning rate is adaptively adjusted based on the relationship between the two factors. The learning rate update formula is: in, η l A fixed learning rate is set for the position filter. η u These are the learning rate update coefficients, used to adjust the learning rate for frame t. η t Let be the learning rate for the updated frame t; Based on the average peak correlation energy (APCE) of frame t and the threshold confidence threshold TH APCE By comparison, the learning rate update coefficients are obtained. η u： Where m is the learning rate update constraint factor; For the target feature template and location filter parameters, a linear weighted update model is used to complete the iteration. The update formula is: in, x t-1 and α t-1 These are the target template and filter parameters from the previous frame, respectively. x new and α new These are the target template and filter parameters calculated for frame t, respectively. x t and α t These are the updated target template and filter parameters for frame t, respectively. η The preset adjustment coefficient; For a scaling filter, its parameters are updated in the frequency domain, and the update formula is: in, and For the first The parameters of the scale filter after frequency domain update and These are the Fourier transforms of the desired Gaussian response and the input features, respectively. The conjugate Fourier transform of the input features. is the learning rate of the scaling filter.
8. A target tracking system based on multi-layer features, used to implement the method as described in any one of claims 1 to 7, characterized in that, The system includes: a data acquisition and initialization unit, a candidate region delineation unit, a deep feature extraction unit, a feature fusion unit, a target localization unit, a scale estimation unit, and an update unit; The data acquisition and initialization unit is used to acquire the target initial position and target initial scale of the video image sequence and the first frame image, and to complete the parameter initialization of the position filter, scale filter and target feature template based on the initial position and initial scale. The candidate region delineation is used to determine candidate regions for the t-th frame image in the video image sequence based on the target position and target scale of the (t-1)-th frame, where t is a positive integer and its initial value is 2; The deep feature extraction unit has a built-in pre-trained VGG-19 network, which is used to input the candidate region into the VGG-19 network, extract the depth features of the first convolutional layer, the second convolutional layer and the third convolutional layer in the network, and calculate the correlation response maps of the depth features of the first convolutional layer, the second convolutional layer and the third convolutional layer based on the position filter, respectively, as the depth feature response maps of each layer. The feature fusion unit is used to calculate the peak-to-sidelobe ratio of the depth feature response map of the three convolutional layers, and then adaptively fuse the feature weights of the three convolutional layers based on the peak-to-sidelobe ratio to output the target localization response map of the t-th frame. The target positioning unit is used to perform confidence analysis on the target positioning response map and determine the target position in frame t based on the analysis results; when the confidence is lower than a preset threshold, the target position in frame t-1 is corrected and the target position in frame t is re-output. The scale estimation unit is used to calculate the target scale estimate in frame t. The updating unit is used to update the parameters of the position filter, the scale filter, and the target feature template based on the target position in the t-th frame and the target scale estimate in the t-th frame. Let t = t + 1, then return to the step of determining the candidate region for the t-th frame of the video image sequence based on the target position and target scale of the (t-1)-th frame, until the target tracking of the video image sequence is completed.
9. A readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, the processor performs the steps of the method as described in any one of claims 1 to 7.
10. A computer device, comprising a memory and a processor, characterized in that, The memory stores a computer program that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 7.