A long-term target tracking method and system based on improved Staple

Abstract

The present invention discloses a long-term target tracking method and system based on improved Staple, firstly, a color filter H is obtained according to the color features and directional gradient histogram features of the target in the current image extracted, and then the color filter H is obtained. color , gradient filter H hog , scale-dependent filter H scale and the minimum mean square error and filter H MOSSE , and then obtain the target position through feature fusion, according to the scale-related filter H scale The target tracking frame with the best scale is obtained, that is, the tracking range is locked, and then it is determined whether the target is occluded. If no occlusion occurs, the color filter H is updated in a linear manner. color , gradient filter H hog and the minimum mean square error and filter H MOSSE , enter the next frame of image; if the target is blocked, use particle filtering to search for the target again and track it, and enter the next frame of image. Then repeat the above process until the last frame of image completes target tracking.

Description

Technical Field

[0001] The present invention relates to the technical field of target tracking, and in particular to a long-term target tracking method and system based on improved Staple. Background Art

[0002] As one of the key research directions in computer vision, visual target tracking technology has been widely used in various fields such as intelligent transportation, human-computer interaction, aerospace, and medical diagnosis. The main task of the tracking algorithm is to automatically give information such as the target position in subsequent frames based on the target selected in the first frame. In actual tracking tasks, affected by complex environments, the algorithm often faces challenges such as motion blur, target rotation changes, and target occlusion. These are all difficult problems that need to be solved in the field of target tracking.

[0003] At present, the algorithms in the field of target tracking can be mainly divided into two categories: those based on correlation filtering and those based on deep learning. Compared with the low frame rate deep learning-based target tracking algorithms, the target tracking algorithms based on correlation filtering have been widely studied by scholars at home and abroad due to their excellent comprehensive performance. The correlation filtering algorithm approximates the expected response by the correlation operation between the filter template and the search area features to learn the correlation filter used to classify the target and background. Since the operation process is mapped to the frequency domain, the algorithm can be continuously detected and updated with very low computational complexity. The more famous algorithms include MOSSE (Minimum Output Sum of Squared Error), CSK, KCF (kernelized correlation filter) and Staple (Sum of Template And Pixel-wise Learners complementary learning model).

[0004] Staple uses a complementary learning method to improve the algorithm's adaptability to scenes. It has an outstanding performance in target tracking methods based on correlation filtering, and continues to attract further research. However, recent studies often enhance the discrimination ability of the Staple algorithm from the perspective of response fusion. Some studies have proposed the use of local sensitivity histograms and fusion methods based on relative confidence, and some studies have proposed adaptive fusion methods based on maximum response peak and average peak correlation energy. However, the original algorithm still has the following problems in long-term tracking tasks: (1) The color feature-based model used by Staple is prone to introduce some background information when extracting target features. (2) When the target is occluded, the Staple model is prone to drift to the point where the target disappears, causing tracking failure. Therefore, it is necessary to improve the Staple algorithm in a targeted manner to improve the stability of the algorithm in long-term tracking tasks. Summary of the invention

[0005] In view of this, the present invention provides a long-term target tracking method and system based on improved Staple, which can track the target for a long time with high stability and solve the problem that the target is prone to tracking failure under long-term occlusion.

[0006] The specific technical solution adopted by the present invention is as follows:

[0007] A long-term target tracking method based on improved Staple, including:

[0008] Step 1: Get the color filter H of the current frame image color , gradient filter H hog , minimum mean square error and filter H MOSSE ;

[0009] Step 2: extract the color features and directional gradient histogram features of the target in the next frame image, fuse the response of the directional gradient histogram feature and the response of the color feature, and obtain the target position according to the fusion result;

[0010] Step 3: The peak value max of the response of the directional gradient histogram feature described in step 2 hog and the peak value of the color feature response max color , combined with the minimum mean square error and filter H described in step 1 MOSSE Perform secondary detection on the target to determine whether the target is blocked;

[0011] Step 4: If the target is not blocked, update the color filter H color , gradient filter H hog and the minimum mean square error and filter H MOSSE , enter the next frame image; if the target is blocked, search for the target again and track it, and enter the next frame image;

[0012] Step 5: Repeat steps 1 to 4 until the target is tracked in the last frame.

[0013] Furthermore, in the step 1, the color filter H is obtained by training according to the color feature of the target in the current frame image. color and the minimum mean square error and filter H MOSSE ; The gradient filter H is obtained according to the directional gradient histogram feature training hog ;

[0014] The scale-dependent filter H is obtained by training the directional gradient histogram features of different scales. scale ; The scale-dependent filter H scale Get the target tracking box of the best scale.

[0015] Furthermore, in the step 2, the color feature used for fusion is the enhanced color feature obtained by mask extraction; the mask extraction method is: firstly obtain the prior probability p(b x |l∈O) and the prior probability p(l∈O) of the target area, and then the pixel space reliability can be obtained according to the Bayesian formula:

[0016]

[0017] Among them, x=l represents a pixel point in the detection area, b x represents the pixel value of the pixel, m is the target model, O is the target area, and B is the background area;

[0018] A reliable target area is obtained according to the pixel space reliability, and then a mask of the target is obtained, and enhanced color features are obtained according to mask extraction.

[0019] Furthermore, in the step 2, the target position is obtained according to the fusion result by weighted fusion of the response of the oriented gradient histogram feature and the response of the enhanced color feature, and the position corresponding to the peak value of the weighted fusion result is used as the target position.

[0020] Furthermore, in step three, the second detection of the target to determine whether the target is blocked is as follows:

[0021] According to the peak value of the response of the directional gradient histogram feature hog and the peak value of the enhanced color feature response max color , calculate and obtain the cross-peak sidelobe ratios CPSR1 and CPSR2; if CPSR1 is less than 1.5 and CPSR2 is less than 3.5, the target is considered to be obscured; otherwise, the target is not obscured.

[0022] Further, in step 4, the re-searching and tracking of the target is:

[0023] Using the particle filtering method, the color characteristics of the pixel points are used to characterize the particle characteristics. According to the color characteristics of each particle, the minimum mean square error and the filter H are combined. MOSSE The weight of each particle is determined, and the state is predicted according to the weight of the particle to obtain the target position and track it.

[0024] A long-term target tracking system based on improved Staple includes: filter training module, feature extraction module, fusion module, judgment module, update module and iteration module;

[0025] The filter training module is used to obtain the color filter H of the current frame image. color , gradient filter Hhog and the minimum mean square error and filter H MOSSE ;

[0026] The feature extraction module is used to extract the color features and directional gradient histogram features of the target in the next frame image;

[0027] The fusion module is used to fuse the response of the oriented gradient histogram feature and the response of the color feature, and obtain the target position according to the fusion result;

[0028] The judgment module is used to determine the peak value of the response of the directional gradient histogram feature of the feature extraction module. hog and the peak value of the enhanced color feature response max color , and the minimum mean square error of the filter training module and the filter H MOSSE Perform secondary detection on the target to determine whether the target is blocked, and send the judgment result to the update module;

[0029] The updating module includes an updating unit and a re-detection unit; the updating unit is used to update the color filter H when the target is not blocked. color , gradient filter H hog and the minimum mean square error and filter H MOSSE , enter the next frame of image; the re-detection unit is used to re-search and track the target when the target is blocked, and enter the next frame of image;

[0030] The iteration module is used to determine whether the target tracking of the last frame image is completed. If not, an iteration instruction is sent to the filter training module to repeat the above process until the target tracking of the last frame image is completed.

[0031] Furthermore, the color features and directional gradient histogram features of the target extracted by the feature extraction module are sent to the filter training module; in the filter training module, the color filter H is obtained according to the color feature training of the target in the current frame image. color and the minimum mean square error and filter H MOSSE ; The gradient filter H is obtained according to the directional gradient histogram feature training hog ;

[0032] The scale-dependent filter H is obtained by training the directional gradient histogram features of different scales. scale ; The scale-dependent filter H scale Get the target tracking box of the best scale.

[0033] Furthermore, in the fusion module, the color feature used for fusion is a color feature enhanced by mask extraction; the mask extraction method is: firstly obtain the prior probability p(b x |l∈O) and the prior probability p(l∈O) of the target area, and then the pixel space reliability can be obtained according to the Bayesian formula:

[0034]

[0035] Among them, x=l represents a pixel point in the detection area, b x represents the pixel value of the pixel, m is the target model, O is the target area, and B is the background area;

[0036] A reliable target area is obtained according to the pixel space reliability, and then a mask of the target is obtained, and enhanced color features are obtained according to mask extraction.

[0037] Furthermore, in the judgment module, the target is detected twice to determine whether the target is blocked: according to the peak value max of the response of the directional gradient histogram feature hog and the peak value of the enhanced color feature response max color , calculate and obtain the cross-peak sidelobe ratios CPSR1 and CPSR2; if CPSR1 is less than 1.5 and CPSR2 is less than 3.5, the target is considered to be obscured; otherwise, the target is not obscured;

[0038] In the re-detection unit, the particle filter is used to re-search and track the target, specifically: the color characteristics of the pixel points are used to represent the particle characteristics, and the color characteristics of each particle are combined with the minimum mean square error and the filter H MOSSE The weight of each particle is determined, and the state is predicted according to the weight of the particle to obtain the target position and track it.

[0039] Beneficial effects:

[0040] (1) A long-term target tracking method based on improved Staple is proposed by training the scale-dependent filter H scale Keeping the target tracking within the optimal range is beneficial to improving the stability of target tracking. Obtaining enhanced color features through mask extraction is beneficial to improving the tracking accuracy in deformation and motion blur scenes. MOSSE The target is detected again to obtain a more accurate occlusion status judgment. When the current target is occluded, the motion estimation method of particle filtering is used to achieve target relocation, thereby improving the tracking success rate when the target is occluded.

[0041] (2) Use particle filtering to search for and track the target again. The color histogram is still used as the particle feature extraction method, so that the minimum mean square error and filter H can be reused. MOSSE , and because of the use of the minimum mean square error and filter H MOSSE Guided particle displacement, generally only 3-5 iterations are needed to obtain accurate re-detection results, which improves the detection efficiency of the target and thus improves the target tracking efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 The present invention is a flowchart of a long-term target tracking method based on improved Staple.

[0043] Figure 2 This is the result diagram of spatial reliability calculation.

[0044] Figure 3 This is the mask extraction result map.

[0045] Figure 4 This is the target response map when mask extraction is not added.

[0046] Figure 5 This is the response map of the target after adding mask extraction.

[0047] Figure 6 This is the response result diagram after secondary detection at the response peak.

[0048] Figure 7 Schematic diagram of using particle filtering to re-detect a target after it is determined to be occluded. DETAILED DESCRIPTION

[0049] A long-term target tracking method based on improved Staple is firstly trained to obtain the color filter H according to the color features and directional gradient histogram features of the target in the current image. color , gradient filter H hog , scale-dependent filter H scale and the minimum mean square error and filter H MOSSE , and then obtain the target position through feature fusion, according to the scale-related filter H scale The target tracking frame with the best scale is obtained, that is, the tracking range is locked, and then it is determined whether the target is occluded. If no occlusion occurs, the color filter H is updated in a linear manner. color , gradient filter H hog and the minimum mean square error and filter H MOSSE , enter the next frame of image; if the target is blocked, use particle filtering to search for the target again and track it, and enter the next frame of image. Then repeat the above process until the last frame of image completes target tracking.

[0050] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0051] like Figure 1 As shown, a long-term target tracking method based on improved Staple of the present invention comprises the following steps:

[0052] Step 1: Get the color filter H of the current frame image color , gradient filter H hog , scale-dependent filter H scale and the minimum mean square error and filter H MOSSE The color filter H is obtained by training the color features of the target in the current frame image. color and the minimum mean square error and filter H MOSSE ; Obtain the gradient filter H based on the directional gradient histogram feature training hog ; The scale-dependent filter H is obtained by training the directional gradient histogram features of different scales scale .

[0053] First, we need to extract the color features and directional gradient histogram features of the target in the current frame image. According to the target selected by the observer in the first frame, we first extract the color features and directional gradient histogram HOG features of the target, and train two independent correlation filters H respectively. color and H hog Among them, G is the expected output, i.e., the Gaussian function, the symbol H is the conjugate transpose, ⊙ is the element-by-element multiplication, ρ is the number of pixels in the region, and λ hog and λ color H hog and H color The regular penalty constant is usually 0.001.

[0054]

[0055]

[0056] Then sample the multi-scale HOG features of the target [F1, F2, ... F d ], train a one-dimensional scale-dependent filter H scale , the training method is the same as H hog Finally, the color histogram feature f of the target is used hist , train a minimum mean square error and MOSSE filter H MOSSE :

[0057]

[0058] Step 2: Extract the color features and directional gradient histogram features of the target in the next frame image, and obtain the enhanced color features through mask extraction; fuse the response of the directional gradient histogram feature and the response of the enhanced color feature to obtain the target position. Figure 1 The position prediction in; at the same time according to the scale correlation filter H scale Get the target tracking box of the best scale.

[0059] First, extract the color features f of the target from the search area in the previous frame image. color and HOG feature f hog , for HOG features, directly calculate its response y hog :

[0060]

[0061] For color features, the spatial prior probability of each pixel position with respect to the target can be expressed by the Ivankonikov kernel function as p(b x |l∈O), let the prior probability of the target area be expressed as:

[0062] p(l∈O)=|O| / (|O|+|B|)

[0063] According to the Bayesian formula, the pixel space reliability can be obtained:

[0064]

[0065] The target reliable area is obtained according to the pixel space reliability, and then the target mask is obtained, and the enhanced color features are obtained according to the mask extraction. The result is truncated to obtain Figure 2 The target reliable area shown in the figure is finally eliminated by using morphological filtering operation to eliminate the smaller interference area, and then we can get Figure 3 The final target mask is shown. Based on the mask, the color features of the target can be re-extracted and calculated. Figure 5 The response y of the color feature is shown color ,contrast Figure 4 The response without masking is shown, and it can be seen that the response range is clearly constrained.

[0066] Finally, the two response maps are weighted fused:

[0067] y=λy color +(1-λ)y hog

[0068] The peak of the fusion result is taken as the target position.

[0069] Step 3: The peak value max of the response based on the directional gradient histogram feature of step 2 hogand the peak value of the enhanced color feature response max color , combining the minimum mean square error of step one and the filter H MOSSE By performing a secondary detection on the target, it is possible to determine whether the target is obscured, corresponding to the response quality assessment in Figure (1).

[0070] The peak value of the response according to the directional gradient histogram feature hog and the peak value of the enhanced color feature response max color , calculate and obtain the cross-peak sidelobe ratios CPSR1 and CPSR2; if CPSR1 is less than 1.5 and CPSR2 is less than 3.5, the target is considered to be obscured; otherwise, the target is not obscured.

[0071] Using H MOSSE Calculate the cross-peak-to-sidelobe ratio:

[0072]

[0073]

[0074] The calculation results using common indicators such as PSR and APCE are as follows Figure 6 As shown, the calculation results using the indicators in this article are as follows Figure 7 As shown, it can be seen that the indicator in this paper is more discriminative of the tracking status of the target.

[0075] PSR is the peak ratio. According to experience, when CPSR1 is less than 1.5 and CPSR2 is less than 3.5, it can be judged that the target is blocked.

[0076] Step 4: According to the target occlusion obtained in step 3, if the target is not occluded, the color filter H is updated in a linear manner. color , gradient filter H hog and the minimum mean square error and filter H MOSSE , enter the next frame image; if the target is blocked, the particle filter is used to search for the target again and track it, and enter the next frame image.

[0077] If the target is judged to be unobstructed, based on the linear update filter, samples of different scales are taken with the predicted position as the center to form a scale pyramid for scale prediction as shown in Figure (1). Assuming the size of the image block where the target is located is M×N, the scale can be preset as follows according to the general law of scale change:

[0078]

[0079] Where a is the preset scale factor, n is the scale coefficient, and s is the scale range. Then, all samples of different sizes are scaled to the original size (M×N), and their j-dimensional fast gradient histogram FHOG features are extracted respectively, and the features of each scale are strung into a one-dimensional vector d=M×N×j. Finally, use The scale response is calculated, and the coefficient at the extreme value of the response is taken. The optimal target frame can be determined by combining the aspect ratio of the target.

[0080] If the target is judged to be blocked, the color histogram features of the pixels are used to characterize the particle features. The color features of each particle are combined with the minimum mean square error and the filter H. MOSSE The weight of each particle is determined, and the state is predicted according to the weight of the particle to obtain the target position and track it.

[0081] The specific steps are:

[0082] (1) Assume that the particle set can be represented as S = {s n ,w n |n=1,...,N}, where s={x,y,v x ,v y} represents the state of the particle, w represents the particle weight, then the cumulative probability of the particle at time t-1 is:

[0083]

[0084] First, use this cumulative probability to generate an accumulation array in the range of [0,1], and then use uniform distribution sampling to obtain a new particle set S′ t-1 .

[0085] (2) Update the particle set S using the set state equation t =AS′ t-1 +Bg t-1 Get the particle set S at the current moment t , where A and B are constant matrices, g t-1 is a Gaussian perturbation.

[0086] (3) For S t Each particle in the collection Its corresponding Extract the candidate area for the center and calculate the distribution histogram, and then use the MOSSE filter to calculate the response for each particle, and the weight of each particle at time t can be obtained as follows:

[0087]

[0088] For the particle set S t The weighted average of all particle states in the re-detection module can be used to obtain the final target position prediction:

[0089]

[0090] Step 5: Repeat steps 1 to 4 until the target is tracked in the last frame.

[0091] According to the above-mentioned long-term target tracking method based on improved Staple, the present invention also provides a long-term target tracking system based on improved Staple, including: a filter training module, a feature extraction module, a fusion module, a judgment module, an update module and an iteration module.

[0092] The filter training module is used to obtain the color filter H of the current frame image. color , gradient filter H hog , scale-dependent filter H scale and the minimum mean square error and filter H MOSSE .

[0093] The feature extraction module is used to extract the color features and directional gradient histogram features of the target in the next frame image.

[0094] The fusion module is used to fuse the response of the oriented gradient histogram feature and the response of the color feature to obtain the target position.

[0095] The judgment module is used to extract the peak value of the response of the directional gradient histogram feature of the feature extraction module. hog and the peak value of the enhanced color feature response max color , and the minimum mean square error of the filter training module and the filter H MOSSE Perform secondary detection on the target to determine whether the target is blocked, and send the judgment result to the update module.

[0096] The update module includes an update unit and a re-detection unit; the update unit is used to update the color filter H in a linear manner when the target is not blocked. color , gradient filter H hog and the minimum mean square error and filter H MOSSE , enter the next frame image; the re-detection unit is used to re-search and track the target when the target is blocked, and enter the next frame image.

[0097] The iteration module is used to determine whether the target tracking of the last frame image is completed. If not, an iteration instruction is sent to the filter training module to repeat the above process until the target tracking of the last frame image is completed.

[0098] The color features and directional gradient histogram features of the target extracted by the feature extraction module are sent to the filter training module; in the filter training module, the color filter H is obtained according to the color features of the target in the current frame image. color and the minimum mean square error and filter H MOSSE ; Obtain the gradient filter H based on the directional gradient histogram feature training hog ; The scale-dependent filter H is obtained by training the directional gradient histogram features of different scales scale .

[0099] The color feature used for fusion is the enhanced color feature obtained by mask extraction. The mask extraction method is: first calculate the prior probability p(b x |l∈O) and the prior probability p(l∈O) of the target area, and then the pixel space reliability can be obtained according to the Bayesian formula:

[0100]

[0101] Among them, x=l represents a pixel point in the detection area, b x represents the pixel value of the pixel, m is the target model, O is the target area, and B is the background area;

[0102] The reliable area of ​​the target is obtained according to the reliability of the pixel space, and then the mask of the target is obtained, and the enhanced color features are obtained according to the mask extraction.

[0103] In the judgment module, the target is detected twice to determine whether the target is blocked: according to the peak value max of the response of the directional gradient histogram feature hog and the peak value of the enhanced color feature response max hist , calculate and obtain the cross-peak sidelobe ratios CPSR1 and CPSR2; if CPSR1 is less than 1.5 and CPSR2 is less than 3.5, the target is considered to be obscured; otherwise, the target is not obscured.

[0104] In the re-detection unit, the particle filter is used to re-search the target and track it as follows: the color features of the pixels are used to represent the particle features, and the color features of each particle are combined with the minimum mean square error and the filter H MOSSE Determine the weight of each particle, predict the state based on the particle weight, and then obtain the target position and track it.

[0105] In each module and unit of the above-mentioned long-term target tracking system based on improved Staple, the specific calculation process involved is the same as the corresponding calculation process in the long-term target tracking method based on improved Staple of the present invention, and will not be described in detail here.

[0106] The above specific embodiments only describe the design principle of the present invention. The shapes and names of the components in the description may be different and are not limited. Therefore, those skilled in the art in the field of the present invention may modify or replace the technical solutions recorded in the above embodiments; and these modifications and replacements do not deviate from the creative purpose and technical solutions of the present invention and should all fall within the protection scope of the present invention.

Claims

1. A long-term target tracking method based on improved Staple, characterized in that: include: Step 1: Get the color filter H of the current frame image color , gradient filter H hog , minimum mean square error and filter H MOSSE ; Step 2: extract the color features and directional gradient histogram features of the target in the next frame image, fuse the response of the directional gradient histogram feature and the response of the color feature, and obtain the target position according to the fusion result; wherein the color feature used for fusion is the enhanced color feature obtained by mask extraction; the mask extraction method is: first calculate the prior probability p(b x |l∈O) and the prior probability p(l∈O) of the target area, and then the pixel space reliability can be obtained according to the Bayesian formula: Among them, x=l represents a pixel point in the detection area, b x represents the pixel value of the pixel, m is the target model, O is the target area, and B is the background area; Obtaining a reliable target region according to the pixel space reliability, thereby obtaining a mask of the target, and obtaining enhanced color features according to mask extraction; The target position is obtained according to the fusion result as follows: weighted fusion is performed on the response of the oriented gradient histogram feature and the response of the enhanced color feature, and the position corresponding to the peak value of the weighted fusion result is used as the target position; Step 3: The peak value max of the response of the directional gradient histogram feature described in step 2 hog and the peak value of the color feature response max color , combined with the minimum mean square error and filter H described in step 1 MOSSE Perform secondary detection on the target to determine whether the target is blocked; Step 4: If the target is not blocked, update the color filter H color , gradient filter H hog and the minimum mean square error and filter H MOSSE , enter the next frame image; if the target is blocked, search for the target again and track it, and enter the next frame image; Step 5: Repeat steps 1 to 4 until the target is tracked in the last frame.

2. The long-term target tracking method according to claim 1, characterized in that: In the step 1, the color filter H is obtained by training according to the color feature of the target in the current frame image. color and the minimum mean square error and filter H MOSSE ; The gradient filter H is obtained according to the directional gradient histogram feature training hog ; The scale-dependent filter H is obtained by training the directional gradient histogram features of different scales. scale ; The scale-dependent filter H scale Get the target tracking box of the best scale.

3. The long-term target tracking method according to claim 1, characterized in that: In step 3, the target is detected again to determine whether the target is blocked: According to the peak value of the response of the directional gradient histogram feature hog and the peak value of the enhanced color feature response max color , calculate and obtain the cross-peak sidelobe ratios CPSR1 and CPSR2; if CPSR1 is less than 1.5 and CPSR2 is less than 3.5, the target is considered to be obscured; otherwise, the target is not obscured.

4. The long-term target tracking method according to claim 1, characterized in that: In step 4, the re-searching and tracking of the target is as follows: Using the particle filtering method, the color characteristics of the pixel points are used to characterize the particle characteristics. The color characteristics of each particle are combined with the minimum mean square error and the filter H MOSSE The weight of each particle is determined, and the state is predicted according to the weight of the particle to obtain the target position and track it.

5. A long-term target tracking system based on improved Staple, characterized in that: include: Filter training module, feature extraction module, fusion module, judgment module, update module and iteration module; The filter training module is used to obtain the color filter H of the current frame image. color , gradient filter H hog and the minimum mean square error and filter H MOSSE ; The feature extraction module is used to extract the color features and directional gradient histogram features of the target in the next frame image; The fusion module is used to fuse the response of the directional gradient histogram feature and the response of the color feature, and obtain the target position according to the fusion result; wherein the color feature used for fusion is the enhanced color feature obtained by mask extraction; the mask extraction method is: firstly obtain the prior probability p(b x |l∈O) and the prior probability p(l∈O) of the target area, and then the pixel space reliability can be obtained according to the Bayesian formula: Among them, x=l represents a pixel point in the detection area, b x represents the pixel value of the pixel, m is the target model, O is the target area, and B is the background area; Obtaining a reliable target region according to the pixel space reliability, thereby obtaining a mask of the target, and obtaining enhanced color features according to mask extraction; The target position is obtained according to the fusion result as follows: weighted fusion is performed on the response of the oriented gradient histogram feature and the response of the enhanced color feature, and the position corresponding to the peak value of the weighted fusion result is used as the target position; The judgment module is used to determine the peak value of the response of the directional gradient histogram feature of the feature extraction module. hog and the peak value of the enhanced color feature response max color , and the minimum mean square error of the filter training module and the filter H MOSSE Perform secondary detection on the target to determine whether the target is blocked, and send the judgment result to the update module; The updating module includes an updating unit and a re-detection unit; the updating unit is used to update the color filter H when the target is not blocked. color , gradient filter H hog and the minimum mean square error and filter H MOSSE , enter the next frame of image; the re-detection unit is used to re-search and track the target when the target is blocked, and enter the next frame of image; The iteration module is used to determine whether the target tracking of the last frame image is completed. If not, an iteration instruction is sent to the filter training module to repeat the above process until the target tracking of the last frame image is completed.

6. The long-term target tracking system according to claim 5, characterized in that: The color features and directional gradient histogram features of the target extracted by the feature extraction module are sent to the filter training module; in the filter training module, the color filter H is obtained according to the color feature training of the target in the current frame image. color and the minimum mean square error and filter H MOSSE ; The gradient filter H is obtained according to the directional gradient histogram feature training hog ; The scale-dependent filter H is obtained by training the directional gradient histogram features of different scales. scale ; The scale-dependent filter H scale Get the target tracking box of the best scale.

7. The long-term target tracking system according to claim 5, characterized in that: In the judgment module, the target is detected twice to determine whether the target is blocked: according to the peak value max of the response of the directional gradient histogram feature hog and the peak value of the enhanced color feature response max color , calculate and obtain the cross-peak sidelobe ratios CPSR1 and CPSR2; if CPSR1 is less than 1.5 and CPSR2 is less than 3.5, the target is considered to be obscured; Otherwise, the target is not occluded; In the re-detection unit, the particle filter is used to re-search and track the target, specifically: the color characteristics of the pixel points are used to represent the particle characteristics, and the color characteristics of each particle are combined with the minimum mean square error and the filter H MOSSE The weight of each particle is determined, and the state is predicted according to the weight of the particle to obtain the target position and track it.

Images