Anti-drone observation tracking system

Abstract

The application discloses an anti-UAV observation and tracking system, comprising the following steps: modeling continuous image frames by using a ViBe algorithm, extracting a foreground mask graph and a dynamic connected region, constructing an initial density kernel model through feature screening, and performing gradient iteration on color density distribution by applying a Mean-Shift tracking algorithm in a target tracking stage, estimating a target observation window position, limiting a search range based on the foreground mask, and adaptively updating the density kernel model when window drift exceeds a threshold value, generating a consistency score and a tracking state identifier through position overlap rate and color density similarity calculation between continuous frames, triggering a recapture mechanism in an abnormal state, and repositioning a target position in combination with current image and historical window information, so that the tracking process stability and accuracy are ensured.

Description

Technical Field

[0001] This invention relates to the field of computer vision and image processing technology, and in particular to an anti-drone observation and tracking system. Background Technology

[0002] With the continuous growth in the number of drones in low-altitude airspace, counter-drone systems are increasingly being used in urban security, patrol and control of important areas, and protection of targets in sensitive areas. At present, most common observation, aiming and tracking systems are based on fixed optoelectronic equipment to acquire images, and combine motion detection, target extraction and tracking algorithms to achieve real-time monitoring of low-altitude moving targets.

[0003] Existing UAV observation and tracking methods still have significant limitations in environments with complex backgrounds, low contrast, and small, fast-moving targets. On the one hand, traditional foreground extraction algorithms based on frame difference or Gaussian mixture modeling are not robust enough in dynamic background interference or weakly textured regions, failing to accurately segment UAV targets and easily causing foreground drift or false detections. On the other hand, the classic Mean-Shift tracking algorithm only relies on color density distribution for target localization, without considering motion information fusion and search range limitations. It is prone to loss when tracking drifts or the target is briefly occluded, and lacks an adaptive recapture mechanism for abnormal tracking states, failing to guarantee continuous and stable tracking.

[0004] Therefore, how to provide an anti-drone observation and tracking system is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an anti-UAV observation and tracking system. This invention integrates background modeling, candidate region screening, color kernel construction, and dynamic correction mechanisms. It describes in detail the processing steps for continuous observation and steady-state tracking of UAV targets in low-altitude airspace, and has the advantages of accurate foreground extraction, strong tracking robustness, and high target drift adaptive correction capability.

[0006] An anti-drone observation and tracking system according to an embodiment of the present invention includes the following steps:

[0007] The image acquisition module is used to acquire continuous low-altitude airspace image frames and generate time-series image data;

[0008] Foreground modeling module, used for... The algorithm models temporal image data and outputs the foreground mask map and dynamically connected regions for each frame;

[0009] The candidate region extraction module is used to receive a set of dynamically connected regions, filter suspicious target regions based on preset size thresholds, velocity estimates and boundary shape indicators, generate color density histograms, and construct an initial density kernel model set.

[0010] The target tracking module receives the initial density kernel model set and the target tracking window of the previous frame, uses the Mean-Shift tracking algorithm to perform gradient iteration on the color density distribution, outputs the current frame observation window and records the tracking trajectory;

[0011] The mask guidance module is used to limit the Mean-Shift search range based on the foreground mask of the current frame, and adaptively update the corresponding kernel model when the window drift exceeds the threshold to generate a corrected observation window;

[0012] The consistency scoring module receives the corrected target window position and density kernel model from two consecutive frames, calculates the score based on the window position overlap rate and model similarity, and outputs the consistency score and tracking status indicator.

[0013] The recapture module is used to roll back the historical window when the tracking status is abnormal, generate candidate regions by combining the current frame foreground mask and relocate the optimal target position, and return to the updated observation window.

[0014] Optionally, modules can be integrated using the following methods:

[0015] Acquire continuous low-altitude airspace image frames and generate time-series image data arranged in chronological order;

[0016] Based on the aforementioned time-series image data, using The algorithm establishes a background model, performs modeling processing on each frame of the image, and outputs the foreground mask map and dynamic connected region set of the corresponding frame.

[0017] Suspicious target regions are selected based on the size parameters, boundary shape, and inter-frame velocity estimation of the dynamically connected regions. Color density histograms are extracted based on the selected regions, and an initial density kernel model corresponding to each target is constructed.

[0018] The initial density kernel model and the target tracking window of the previous frame are used as inputs. The Mean-Shift tracking algorithm is used to perform gradient iteration on the color density distribution and outputs the target observation window of the current frame.

[0019] The foreground mask of the current frame is input together with the target observation window to limit the Mean-Shift search range. When the observation window drifts more than the threshold compared with the previous frame, the density kernel model of the corresponding target is updated to generate the corrected target window.

[0020] The position overlap rate and density similarity are calculated for the target window and density kernel model after correction for two consecutive frames, a consistency score is generated and the corresponding tracking status label is output;

[0021] When the tracking status indicator indicates an anomaly, the system reverts to the historical window with the best recent consistency score, generates candidate regions by combining the current frame foreground mask map, relocates the optimal target position through similarity calculation, and updates the target observation window.

[0022] Optionally, the time-series image data consists of a sequence of image frames continuously acquired at preset time intervals. Each image frame includes a two-dimensional pixel matrix, a frame number index, a timestamp, imaging exposure parameters, a camera internal parameter matrix, and external attitude information.

[0023] Optionally, the foreground mask map and dynamically connected region set of the current frame image may be output as follows:

[0024] The foreground modeling module receives the current frame image from the time-series image data and constructs a pixel background sample set for each pixel position in the current frame image. The pixel background sample set consists of pixel values ​​obtained by sampling the corresponding pixel position in multiple historical frame images.

[0025] based on The algorithm performs a foreground determination operation. For each pixel in the current frame image, it calculates the Euclidean distance between the pixel value and all sample values ​​in the corresponding pixel background sample set, and counts the number of samples whose Euclidean distance is less than the matching threshold. If the number of samples is less than the minimum matching number, the corresponding pixel is determined to be a foreground pixel; otherwise, it is determined to be a background pixel.

[0026] Based on the above pixel foreground determination results, a foreground mask image is generated. The foreground mask image is a binary image, with foreground pixels marked as 1 and background pixels marked as 0.

[0027] Perform connectivity clustering based on 8-neighborhood structure on all foreground pixels in the foreground mask map of the current frame image, and output the dynamic connected region set of the current frame image;

[0028] The background model update operation is performed based on the ViBe algorithm. For pixels in the current frame image that are determined to be background, a sample position is randomly selected from the corresponding pixel background sample set and replaced with the current value of the corresponding pixel. Then, a pixel position is randomly selected in the spatial neighborhood of the corresponding pixel and the current value is written into its pixel background sample set.

[0029] Erosion and dilation operations are performed on the foreground mask image, and the foreground mask image and the set of dynamically connected regions are used as the modeling output of the current frame image.

[0030] Optionally, constructing the initial density kernel model corresponding to each objective includes:

[0031] The candidate region extraction module receives a set of dynamically connected regions and calculates the number of pixels, aspect ratio, inter-frame centroid displacement amplitude, and color channel mean and variance for each spatial region in the set of dynamically connected regions.

[0032] The candidate region extraction module constructs a joint screening condition function for each spatial region based on size threshold, aspect ratio range, minimum speed threshold and color fluctuation threshold, and performs condition function judgment operation on the set of dynamically connected regions to eliminate spatial regions that do not meet all screening conditions and retain them as a set of candidate target regions.

[0033] For each spatial region in the candidate target region set, perform three-channel joint modeling in the RGB color space, construct a combination vector from the red, green, and blue channel values ​​of all pixels in the region, and perform frequency statistics according to discrete intervals to construct a three-dimensional color joint density histogram.

[0034] Normalize the three-dimensional color joint density histogram to obtain the normalized color joint density kernel function, and use the corresponding color joint density kernel function as the initial density kernel model;

[0035] The initial density kernel model is bound to the spatial center point coordinates of the corresponding candidate target region to form the initial target structure unit, which serves as the initial input for the target tracking module.

[0036] Optionally, the target observation window for outputting the current frame image includes:

[0037] Before executing the Mean-Shift tracking algorithm, based on The algorithm generates a foreground mask image and constructs a pixel probability density distribution function for the pixel positions marked as foreground in the mask image;

[0038] The above probability density distribution function is converted into a kernel function to initialize the weight distribution. The initial influence area of ​​the kernel function is set only within the foreground region to generate the foreground constraint initialization window.

[0039] The foreground constraint initialization window is set as the initial observation window. Based on the color density kernel model corresponding to the target region in the previous frame, three-dimensional joint color space modeling is performed to generate the color histogram of the region to be matched.

[0040] Construct a joint similarity kernel function, fusing the color distribution histogram and channel feature vectors to form a dual-channel feature response. Define the similarity kernel function as follows:

[0041] ;

[0042] in, and These are the weighting coefficients for the color distribution channel and the motion feature channel, respectively, satisfying... , express Three-channel joint density histogram Gradient feature descriptors representing channel fusion;

[0043] Perform gradient ascent iteration on the local region centered on the current window in the current frame image, and update the target observation window position based on the direction of the maximum value of the kernel response function until the stopping threshold condition is met;

[0044] Output the updated target observation window and record the center coordinates and response function value of the corresponding window.

[0045] Optionally, generating the corrected target observation window includes:

[0046] The mask guidance module receives the foreground mask image corresponding to the current frame image and the target observation window coordinate information output in the previous frame image. It takes all the foreground pixel positions in the foreground mask image of the current frame image as candidate search regions and constructs a search region limiting mask by combining the position parameters of the target observation window in the previous frame image. The search region limiting mask is used to set the window search boundary in the Mean-Shift tracking algorithm.

[0047] Within the range specified by the search area defined by the mask, the gradient direction of the pixel probability density function is estimated based on the density kernel model of the corresponding target region in the current frame image. The Mean-Shift tracking algorithm is used to perform position iteration and output the initial target observation window position in the current frame image.

[0048] Perform Euclidean distance calculation between the spatial center point of the initial target observation window in the current frame image and the spatial center point of the target observation window in the previous frame image. If the obtained distance value is greater than the preset drift judgment threshold, it is determined that the target window has drifted.

[0049] When a target window drift event occurs, the mask guidance module receives the three-dimensional color values ​​of all pixels in the initial target observation window of the current frame image, reconstructs the density kernel model of the current frame image based on the normalized color joint density histogram, and weights and fuses the density kernel model with the density kernel model of the corresponding target in the previous frame image according to the weight ratio to obtain the updated density kernel model.

[0050] Based on the updated density kernel model, the Mean-Shift tracking algorithm is executed again on the pixels within the mask of the search region in the current frame image, and the updated target observation window coordinates are output.

[0051] Optionally, a consistency score is performed between the corrected target observation window and the density kernel model for two consecutive frames, including:

[0052] The consistency scoring module receives the corrected target observation window coordinate information from the previous frame image and the current frame image, calculates the area of ​​each window based on the coordinates of the upper left and lower right corners of the two windows, and constructs the position overlap rate based on the area of ​​the intersection region and the area of ​​the union region. The position overlap rate is the ratio of the area of ​​the intersection region to the area of ​​the union region.

[0053] The consistency scoring module receives the target density kernel model in the corresponding frame image, uses the normalized three-dimensional color joint density histogram in each model as input features, constructs a color distribution vector space, performs cosine angle calculation accordingly, and outputs the consistency score.

[0054] The consistency score is compared with a preset confidence threshold. When the consistency score is less than the low confidence threshold, a failure flag is output. When the consistency score is between the low confidence threshold and the high confidence threshold, an unstable flag is output. When the consistency score is higher than the high confidence threshold, a normal flag is output, and a tracking status flag is generated.

[0055] Optionally, updating the target observation window and relocating the optimal target position includes:

[0056] The recapture module receives the tracking status flag. When the tracking status flag is a failure flag or an unstable flag, it determines that the target observation window in the current frame image is in an abnormal state.

[0057] The recapture module reverts to the historical observation window with the highest consistency score in the most recent consecutive frames, and combines the foreground mask map of the current frame image to perform candidate region generation operations in the spatial region where the historical observation window is located and its surrounding extended region;

[0058] Spatial aggregation based on connectivity clustering is performed on the region covered by the foreground mask map in the current frame image to extract multiple candidate regions, and pre-screening is performed based on the location parameters, size parameters and spatial proximity of the historical observation window of each candidate region;

[0059] For all pre-screened candidate regions, calculate the corresponding joint color density histogram and perform cosine similarity calculation with the color distribution vector after normalization of the density kernel model in the historical observation window;

[0060] The candidate region corresponding to the maximum value in the cosine similarity calculation is taken as the optimal target position in the current frame image, and the spatial center point of the corresponding region is taken as the updated target observation window.

[0061] The beneficial effects of this invention are:

[0062] (1) By combining the ViBe background modeling algorithm and the Mean-Shift density tracking algorithm, this invention constructs a highly robust continuous observation and aiming method for UAV targets, which realizes high-precision identification and multi-frame steady-state tracking of low-altitude dynamic targets. It can effectively adapt to challenges such as complex background interference, rapid movement of UAVs and scale changes, and significantly improve the timeliness and stability of foreground extraction and target tracking.

[0063] (2) This invention limits the Mean-Shift search range by using a foreground mask map, combines a dynamic kernel model weight fusion update mechanism to adaptively handle the tracking drift problem, and uses a consistency scoring module to calculate the tracking status identifier based on the position overlap rate and density histogram similarity to achieve dynamic evaluation of tracking stability and effectiveness, effectively improving the target recapture capability and system continuous observation capability in occlusion, loss or sudden change scenarios.

[0064] (3) This invention enhances the target representation model’s ability to perceive texture, shape and motion features by constructing candidate region screening functions, RGB color joint modeling and multi-channel feature fusion, and improves the Mean-Shift algorithm migration update process by combining gradient ascent optimization strategy, so that the tracking window has stronger local positioning ability and fine adjustment ability, thereby improving the practicality and engineering deployment reliability of the UAV tracking system. Attached Figure Description

[0065] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0066] Figure 1 This is a flowchart of an anti-drone observation and tracking system proposed in this invention;

[0067] Figure 2 This is a flowchart of the target observation window update process in this invention, which integrates channel features and performs Mean-Shift gradient iteration. Detailed Implementation

[0068] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0069] refer to Figures 1-2 An anti-drone observation and tracking system includes the following steps:

[0070] The image acquisition module is used to acquire continuous low-altitude airspace image frames and generate time-series image data;

[0071] Foreground modeling module, used for... The algorithm models temporal image data and outputs the foreground mask map and dynamically connected regions for each frame;

[0072] The candidate region extraction module is used to receive a set of dynamically connected regions, filter suspicious target regions based on preset size thresholds, velocity estimates and boundary shape indicators, generate color density histograms, and construct an initial density kernel model set.

[0073] The target tracking module receives the initial density kernel model set and the target tracking window of the previous frame, uses the Mean-Shift tracking algorithm to perform gradient iteration on the color density distribution, outputs the current frame observation window and records the tracking trajectory;

[0074] The mask guidance module is used to limit the Mean-Shift search range based on the foreground mask of the current frame, and adaptively update the corresponding kernel model when the window drift exceeds the threshold to generate a corrected observation window;

[0075] The consistency scoring module receives the corrected target window position and density kernel model from two consecutive frames, calculates the score based on the window position overlap rate and model similarity, and outputs the consistency score and tracking status indicator.

[0076] The recapture module is used to roll back the historical window when the tracking status is abnormal, generate candidate regions by combining the current frame foreground mask and relocate the optimal target position, and return to the updated observation window.

[0077] In this embodiment, the modules are interconnected using the following method:

[0078] Acquire continuous low-altitude airspace image frames and generate time-series image data arranged in chronological order;

[0079] Based on the aforementioned time-series image data, using The algorithm establishes a background model, performs modeling processing on each frame of the image, and outputs the foreground mask map and dynamic connected region set of the corresponding frame.

[0080] Suspicious target regions are selected based on the size parameters, boundary shape, and inter-frame velocity estimation of the dynamically connected regions. Color density histograms are extracted based on the selected regions, and an initial density kernel model corresponding to each target is constructed.

[0081] The initial density kernel model and the target tracking window of the previous frame are used as inputs. The Mean-Shift tracking algorithm is used to perform gradient iteration on the color density distribution and outputs the target observation window of the current frame.

[0082] The foreground mask of the current frame is input together with the target observation window to limit the Mean-Shift search range. When the observation window drifts more than the threshold compared with the previous frame, the density kernel model of the corresponding target is updated to generate the corrected target window.

[0083] The position overlap rate and density similarity are calculated for the target window and density kernel model after correction for two consecutive frames, a consistency score is generated and the corresponding tracking status label is output;

[0084] When the tracking status indicator indicates an anomaly, the system reverts to the historical window with the best recent consistency score, generates candidate regions by combining the current frame foreground mask map, relocates the optimal target position through similarity calculation, and updates the target observation window.

[0085] In this embodiment, the time-series image data consists of a sequence of image frames continuously acquired at preset time intervals. Each image frame includes a two-dimensional pixel matrix, a frame number index, a timestamp, imaging exposure parameters, a camera internal parameter matrix, and external attitude information.

[0086] In this embodiment, the foreground mask map and dynamically connected region set of the current frame image are output as follows:

[0087] The foreground modeling module receives the current frame image from the time-series image data and constructs a pixel background sample set for each pixel position in the current frame image. The pixel background sample set consists of pixel values ​​obtained by sampling the corresponding pixel position in multiple historical frame images.

[0088] based on The algorithm performs a foreground determination operation. For each pixel in the current frame image, it calculates the Euclidean distance between the pixel value and all sample values ​​in the corresponding pixel background sample set, and counts the number of samples whose Euclidean distance is less than the matching threshold. If the number of samples is less than the minimum matching number, the corresponding pixel is determined to be a foreground pixel; otherwise, it is determined to be a background pixel.

[0089] Based on the above pixel foreground determination results, a foreground mask image is generated. The foreground mask image is a binary image, with foreground pixels marked as 1 and background pixels marked as 0.

[0090] Perform connectivity clustering based on 8-neighborhood structure on all foreground pixels in the foreground mask map of the current frame image, and output the dynamic connected region set of the current frame image. The dynamic connected region set consists of multiple spatially continuous foreground pixel regions.

[0091] The background model update operation is performed based on the ViBe algorithm. For pixels in the current frame image that are determined to be background, a sample position is randomly selected from the corresponding pixel background sample set and replaced with the current value of the corresponding pixel. Then, a pixel position is randomly selected in the spatial neighborhood of the corresponding pixel and the current value is written into its pixel background sample set.

[0092] Erosion and dilation operations are performed on the foreground mask image, and the foreground mask image and the set of dynamically connected regions are used as the modeling output of the current frame image;

[0093] The minimum number of matches is The fixed integer threshold used for foreground determination in the algorithm limits the minimum number of similar samples that a pixel value in the current frame image must satisfy in its pixel background sample set. The Euclidean distance between the current pixel value and the corresponding pixel position sample values ​​collected in the historical frames is calculated one by one. When the number of samples with a distance less than the set threshold is not less than the minimum matching number, the pixel is determined to be the background; otherwise, it is determined to be the foreground. The minimum matching number is usually set to 2, which is a parameter determined empirically based on probability statistics and misjudgment tolerance mechanism to ensure the sensitivity of the foreground response and the robustness of the background modeling.

[0094] In this embodiment, constructing the initial density kernel model corresponding to each target includes:

[0095] The candidate region extraction module receives a set of dynamically connected regions and calculates the number of pixels, aspect ratio, inter-frame centroid displacement amplitude, and color channel mean and variance for each spatial region in the set of dynamically connected regions.

[0096] The candidate region extraction module constructs a joint screening condition function for each spatial region based on size threshold, aspect ratio range, minimum speed threshold and color fluctuation threshold, and performs condition function judgment operation on the set of dynamically connected regions to eliminate spatial regions that do not meet all screening conditions and retain them as a set of candidate target regions.

[0097] For each spatial region in the candidate target region set, perform three-channel joint modeling in the RGB color space, construct a combination vector from the red, green, and blue channel values ​​of all pixels in the region, and perform frequency statistics according to discrete intervals to construct a three-dimensional color joint density histogram.

[0098] Normalize the three-dimensional color joint density histogram to obtain the normalized color joint density kernel function, and use the corresponding color joint density kernel function as the initial density kernel model;

[0099] The initial density kernel model is bound to the spatial center point coordinates of the corresponding candidate target region to form an initial target structure unit, which is then used as the initial input for the target tracking module.

[0100] The conditional function judgment operation refers to the candidate region extraction module sequentially calculating the corresponding structural attribute parameters for each dynamically connected region, including the number of pixels, boundary aspect ratio, inter-frame centroid displacement magnitude, and color channel variance. These parameters are then input into a combinational logic function for joint judgment. The combinational logic function connects the judgment conditions using a logical "AND" relationship. Only when all parameters satisfy the corresponding numerical range constraints is the dynamically connected region determined as a candidate target region; otherwise, it is discarded. For example, the number of pixels must be greater than a minimum size threshold, the aspect ratio must be within a set range, the centroid displacement must exceed a minimum motion threshold, and the color variance must be higher than a minimum texture change threshold. Only when all conditions are met is the conditional function judgment result true, and the region is retained.

[0101] In this embodiment, the target observation window that outputs the current frame image includes:

[0102] Before executing the Mean-Shift tracking algorithm, based on The algorithm generates a foreground mask image and constructs a pixel probability density distribution function for the pixel positions marked as foreground in the mask image;

[0103] The above probability density distribution function is converted into a kernel function to initialize the weight distribution. The initial influence area of ​​the kernel function is set only within the foreground region to generate the foreground constraint initialization window.

[0104] The foreground constraint initialization window is set as the initial observation window. Based on the color density kernel model corresponding to the target region in the previous frame, three-dimensional joint color space modeling is performed to generate the color histogram of the region to be matched.

[0105] Construct a joint similarity kernel function, fusing the color distribution histogram and channel feature vectors to form a dual-channel feature response. Define the similarity kernel function as follows:

[0106] ;

[0107] in, and These are the weighting coefficients for the color distribution channel and the motion feature channel, respectively, satisfying... , express Three-channel joint density histogram Gradient feature descriptors representing channel fusion;

[0108] This formula is based on feature fusion strategies in the fields of image processing and pattern recognition, and its initial source is the target modeling density function in the Mean-Shift algorithm, where... Represents pixels of The response values ​​of the three-channel joint density histogram are derived from the kernel density estimation method. This represents the dual-channel feature response function extracted jointly for horizontal and vertical gradients within the current window region, constructed as an improved gradient direction histogram feature. This is achieved by constructing a fused feature response function. Introducing adjustable weight parameters , Controlling the ratio of color density to the response of the gradient structure, where Both are dimensionless real numbers, used to adaptively balance color and shape features in target drift search. This structure introduces structural gradient factors on the basis of density kernel function, which substantially improves the stability of Mean-Shift algorithm in color ambiguity and target occlusion scenarios, and constitutes an innovative path of fusion modeling at the algorithm level.

[0109] Perform gradient ascent iteration on the local region centered on the current window in the current frame image, and update the target observation window position based on the direction of the maximum value of the kernel response function until the stopping threshold condition is met;

[0110] Output the updated target observation window and record the center coordinates and response function values ​​of the corresponding window;

[0111] The pixel probability density distribution function includes, for all pixels within the target observation window, firstly extracting the values ​​of the red, green, and blue channels of each pixel in the RGB color space, and representing each pixel as a three-dimensional color vector. The RGB color space is divided into several discrete color intervals according to a preset color resolution. Interval mapping is performed on all pixel vectors within the observation window, the number of pixels in each color interval is counted, and the relative frequency of the color interval is obtained by dividing the number of pixels in each color interval by the total number of pixels, forming a normalized three-dimensional histogram. This three-dimensional histogram constitutes the pixel probability density distribution function corresponding to the window, and each element represents the probability of occurrence of the corresponding color interval in the target area.

[0112] The execution of the 3D joint color space modeling includes extracting the values ​​of each pixel in the RGB three color channels for all pixels in the candidate target region, forming the following structure: A three-dimensional color vector. Color values ​​are uniformly quantized according to set color precision parameters, for example, each channel is divided into... If there are intervals, then the entire color space is divided into 10 intervals. A cubic block, for each pixel, according to its The quantization interval of each component is mapped to its corresponding color cube. The occurrence frequency of all pixels within the candidate region in each color cube is counted, forming a three-dimensional color joint frequency tensor. This tensor is then normalized to ensure its summation. Generate a normalized density histogram of the candidate region in the three-dimensional joint color space;

[0113] The gradient ascent iteration includes using the target observation window from the previous frame in the current frame as the initial window, constructing a normalized 3D color joint density kernel function based on the pixels within the initial window, calculating the similarity distribution between this kernel function and pixels in the candidate region, traversing each pixel in the candidate region, constructing a sliding sub-window of the same size as the current window centered on that pixel, calculating the similarity response value between the color distribution within the sliding sub-window and the target density kernel function, and weighting the center positions of all sub-windows according to their corresponding similarities to obtain the new window center coordinates. The observation window is then translated to the new window center coordinates, and the above process is repeated until the change in the window center position is less than a preset iteration convergence threshold, or the maximum number of iterations is reached.

[0114] In this embodiment, generating the corrected target observation window includes:

[0115] The mask guidance module receives the foreground mask image corresponding to the current frame image and the target observation window coordinate information output in the previous frame image. It takes all the foreground pixel positions in the foreground mask image of the current frame image as candidate search regions and constructs a search region limiting mask by combining the position parameters of the target observation window in the previous frame image. The search region limiting mask is used to set the window search boundary in the Mean-Shift tracking algorithm.

[0116] Within the range specified by the search area defined by the mask, the gradient direction of the pixel probability density function is estimated based on the density kernel model of the corresponding target region in the current frame image. The Mean-Shift tracking algorithm is used to perform position iteration and output the initial target observation window position in the current frame image.

[0117] Perform Euclidean distance calculation between the spatial center point of the initial target observation window in the current frame image and the spatial center point of the target observation window in the previous frame image. If the obtained distance value is greater than the preset drift judgment threshold, it is determined that the target window has drifted.

[0118] When a target window drift event occurs, the mask guidance module receives the three-dimensional color values ​​of all pixels in the initial target observation window of the current frame image, reconstructs the density kernel model of the current frame image based on the normalized color joint density histogram, and weights and fuses the density kernel model with the density kernel model of the corresponding target in the previous frame image according to the weight ratio to obtain the updated density kernel model.

[0119] Based on the updated density kernel model, the Mean-Shift tracking algorithm is executed again on the pixels within the mask of the search region in the current frame image, and the updated target observation window coordinates are output.

[0120] In this embodiment, the consistency scoring of the target observation window and density kernel model after correction of two consecutive frames of images includes:

[0121] The consistency scoring module receives the corrected target observation window coordinate information from the previous frame image and the current frame image, calculates the area of ​​each window based on the coordinates of the upper left and lower right corners of the two windows, and constructs the position overlap rate based on the area of ​​the intersection region and the area of ​​the union region. The position overlap rate is the ratio of the area of ​​the intersection region to the area of ​​the union region.

[0122] The consistency scoring module receives the target density kernel model in the corresponding frame image, uses the normalized three-dimensional color joint density histogram in each model as input features, constructs a color distribution vector space, performs cosine angle calculation accordingly, and outputs the consistency score.

[0123] The consistency score is compared with a preset confidence threshold. When the consistency score is less than the low confidence threshold, a failure flag is output. When the consistency score is between the low confidence threshold and the high confidence threshold, an unstable flag is output. When the consistency score is higher than the high confidence threshold, a normal flag is output, and a tracking status flag is generated.

[0124] In this embodiment, updating the target observation window and relocating the optimal target position includes:

[0125] The recapture module receives the tracking status flag. When the tracking status flag is a failure flag or an unstable flag, it determines that the target observation window in the current frame image is in an abnormal state.

[0126] The recapture module reverts to the historical observation window with the highest consistency score in the most recent consecutive frames, and combines the foreground mask map of the current frame image to perform candidate region generation operations in the spatial region where the historical observation window is located and its surrounding extended region;

[0127] Spatial aggregation based on connectivity clustering is performed on the region covered by the foreground mask map in the current frame image to extract multiple candidate regions, and pre-screening is performed based on the location parameters, size parameters and spatial proximity of the historical observation window of each candidate region;

[0128] For all pre-screened candidate regions, calculate the corresponding joint color density histogram and perform cosine similarity calculation with the color distribution vector after normalization of the density kernel model in the historical observation window;

[0129] The candidate region corresponding to the maximum value in the cosine similarity calculation is taken as the optimal target position in the current frame image, and the spatial center point of the corresponding region is taken as the updated target observation window.

[0130] Example 1:

[0131] To verify the feasibility of this invention in practice, it was applied to a field deployment experiment of a certain anti-drone mission system. A high-definition monitoring camera deployed on a ground tripod was used to continuously observe and capture images. The test period covered three lighting conditions: morning, noon, and evening. The target was a multi-rotor drone platform with a flight altitude not exceeding 50 meters and a radius not exceeding 0.5 meters. The flight speed variation range was controlled between 1.5 m / s and 6.8 m / s.

[0132] The image acquisition module acquires video frame data at 100ms intervals, with each frame measuring 1280×720 pixels and a sampling frequency of 10Hz. The generated image frames are arranged by timestamps to form time-series image data. The foreground modeling module constructs a pixel-background sample set based on the ViBe algorithm, performs Euclidean distance determination on each frame's pixels and historical pixel samples, generates a foreground mask map, and then performs 8-neighbor clustering to obtain a set of dynamically connected regions. During foreground modeling, the matching threshold is set to 20, and the minimum matching number is set to 2, effectively distinguishing UAV pixels from background interference.

[0133] The candidate region extraction module calculates the number of pixels, aspect ratio, centroid velocity, and color variance for each connected region. It sets the following criteria: size greater than 900 pixels, aspect ratio between 0.3 and 3.0, velocity not less than 1.2 pixels / frame, and color variance greater than 3.5. Based on the conditional function, it eliminates false target regions and generates a set of candidate target regions.

[0134] The target region is jointly modeled in three channels in the RGB color space to construct a three-dimensional color joint density histogram. After normalization, the initial density kernel model of the target is formed. This model, together with the target space center coordinates, is input into the target tracking module to form the initial target structural unit.

[0135] The target tracking module uses the observation window of the previous frame as its initial position and executes the Mean-Shift tracking algorithm in the current frame image based on the fusion formula of the color distribution histogram and gradient features. The parameters are set as follows: α=0.7, β=0.3, maximum number of iterations is 15, and convergence threshold is 2 pixels. The fused features consist of a color density histogram and a gradient structure vector, which greatly improves tracking stability under occlusion and blurring conditions.

[0136] The masking guidance module further limits the target search to the area marked by the foreground mask image in the current frame, and sets a boundary within 40 pixels centered on the observation window of the previous frame to limit the search range and ensure tracking efficiency. If the Euclidean distance between the target observation window in the current frame and the previous frame exceeds 10 pixels, a drift update mechanism is triggered. The system re-acquires the color information within the current window, updates the normalized color density histogram, and then performs a weighted fusion with the density kernel model of the previous frame at a weight ratio of 0.6:0.4 before performing Mean-Shift iteration again, outputting the corrected observation window.

[0137] The consistency scoring module receives the observation window coordinates and color density kernel model from two consecutive frames of images, calculates the position overlap rate and the cosine similarity of the density kernel model respectively. When the position overlap rate is less than 0.4 or the cosine similarity is less than 0.6, it is marked as abnormal, and the system outputs a tracking status indicator of "failure" or "unstable".

[0138] When the system detects an "unstable" state for two consecutive frames, the recapture module triggers a backoff mechanism, retrieves the historical observation window with the highest score in the previous time period as a reference, reconstructs the candidate region by clustering the mask image within the window and its surrounding 50-pixel radius, and calculates the color vector similarity with the historical density kernel model. The one with the highest similarity is selected as the optimal target position for the current frame, and the target observation window is updated.

[0139] To verify the system's performance, the test team conducted 120 drone trajectory experiments with different paths and speeds in the aforementioned scenario. The system recorded the number of target loss instances, target relocation success rate, average tracking stability score, and average tracking error. The experimental results are shown below:

[0140] Table 1 Target tracking performance evaluation data (total sample size: 120)

[0141]

[0142] Based on the experimental data listed in Table 1, the anti-UAV observation and tracking system proposed in this invention significantly outperforms the traditional Mean-Shift algorithm in several key performance indicators. In terms of average relocation success rate, this system achieves 97.5%, an improvement of 15.2 percentage points compared to the traditional method, indicating higher recovery capability and accuracy assurance after target tracking failure. Regarding the average tracking error, the system's error is controlled within 4.6 pixels, while the traditional algorithm's error is 11.2 pixels, a reduction of nearly 60%, fully verifying the significant optimization effect of the Mean-Shift iterative method, which integrates color density and gradient features, on target positioning accuracy. In terms of tracking stability, the system's average consistency score is 0.89, significantly better than the traditional method's 0.65, indicating stronger consistency maintenance capability in dynamic target continuity assessment and window update strategy. The system only experienced 2 failure events during the entire test, while the traditional algorithm experienced 15, further reflecting the significant improvement in the system's robustness and adaptability to complex environments.

[0143] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A counter-drone observation, aiming, and tracking system, characterized in that, include: The image acquisition module is used to acquire continuous low-altitude airspace image frames and generate time-series image data; The foreground modeling module is used to model temporal image data based on the viBe algorithm, and outputs the foreground mask map and dynamically connected regions for each frame; The candidate region extraction module is used to receive a set of dynamically connected regions, filter suspicious target regions based on preset size thresholds, velocity estimates and boundary shape indicators, generate color density histograms, and construct an initial density kernel model set. The target tracking module receives the initial density kernel model set and the target tracking window of the previous frame, uses the Mean-Shift tracking algorithm to perform gradient iteration on the color density distribution, outputs the current frame observation window and records the tracking trajectory; The mask guidance module is used to limit the Mean-Shift search range based on the foreground mask of the current frame, and adaptively update the corresponding kernel model when the window drift exceeds the threshold to generate a corrected observation window; The consistency scoring module receives the corrected target window position and density kernel model from two consecutive frames, calculates the score based on the window position overlap rate and model similarity, and outputs the consistency score and tracking status indicator. The recapture module is used to backtrack the historical window when the tracking status is abnormal, generate candidate regions by combining the current frame foreground mask and relocate the optimal target position, and return to the updated observation window; The candidate region extraction module includes: The candidate region extraction module receives a set of dynamically connected regions and calculates the number of pixels, aspect ratio, inter-frame centroid displacement amplitude, and color channel mean and variance for each spatial region in the set of dynamically connected regions. The candidate region extraction module constructs a joint filtering condition function for each spatial region based on size thresholds, aspect ratio ranges, minimum velocity thresholds, and color fluctuation thresholds. It then performs a condition function judgment operation on the dynamically connected region set, eliminating regions that are not... Spatial regions that meet all filtering criteria are retained as a set of candidate target regions. For each spatial region in the candidate target region set, perform three-channel joint modeling in the RGB color space, construct a combination vector from the red, green, and blue channel values ​​of all pixels in the region, and perform frequency statistics according to discrete intervals to construct a three-dimensional color joint density histogram. Normalize the three-dimensional color joint density histogram to obtain the normalized color joint density kernel function, and use the corresponding color joint density kernel function as the initial density kernel model; The initial density kernel model is bound to the spatial center point coordinates of the corresponding candidate target region to form an initial target structure unit, which is then used as the initial input for the target tracking module. The masking module includes: The mask guidance module receives the foreground mask image corresponding to the current frame image and the target observation window coordinate information output in the previous frame image. It takes all the foreground pixel positions in the foreground mask image of the current frame image as candidate search regions and constructs a search region limiting mask by combining the position parameters of the target observation window in the previous frame image. The search region limiting mask is used to set the window search boundary in the Mean-Shift tracking algorithm. Within the range specified by the search area defined by the mask, the gradient direction of the pixel probability density function is estimated based on the density kernel model of the corresponding target region in the current frame image. The Mean-Shift tracking algorithm is used to perform position iteration and output the initial target observation window position in the current frame image. Perform Euclidean distance calculation between the spatial center point of the initial target observation window in the current frame image and the spatial center point of the target observation window in the previous frame image. If the obtained distance value is greater than the preset drift judgment threshold, it is determined that the target window has drifted. When a target window drift event occurs, the mask guidance module receives the three-dimensional color values ​​of all pixels in the initial target observation window of the current frame image, reconstructs the density kernel model of the current frame image based on the normalized color joint density histogram, and weights and fuses the density kernel model with the density kernel model of the corresponding target in the previous frame image according to the weight ratio to obtain the updated density kernel model. Based on the updated density kernel model, the Mean-Shift tracking algorithm is executed again on the pixels within the mask of the search region in the current frame image, and the updated target observation window coordinates are output.

2. The anti-drone observation and tracking system according to claim 1, characterized in that, The time-series image data consists of a sequence of image frames continuously acquired at preset time intervals. Each image frame includes a two-dimensional pixel matrix, a frame number index, a timestamp, imaging exposure parameters, a camera internal parameter matrix, and external attitude information.

3. The anti-drone observation and tracking system according to claim 1, characterized in that, The foreground modeling module includes: The foreground modeling module receives the current frame image from the time-series image data and constructs a pixel background sample set for each pixel position in the current frame image. The pixel background sample set consists of pixel values ​​obtained by sampling the corresponding pixel position in multiple historical frame images. The foreground determination operation is performed based on the viBe algorithm. For each pixel in the current frame image, the Euclidean distance between the pixel value and all sample values ​​in the corresponding pixel background sample set is calculated, and the number of samples whose Euclidean distance is less than the matching threshold is counted. If the number of samples is less than the minimum matching number, the corresponding pixel is determined to be a foreground pixel; otherwise, it is determined to be a background pixel. A foreground mask is generated based on the pixel foreground determination result. The foreground mask is a binary image, with foreground pixels marked as 1 and background pixels marked as 0. Perform connectivity clustering based on 8-neighborhood structure on all foreground pixels in the foreground mask map of the current frame image, and output the dynamic connected region set of the current frame image; The background model update operation is performed based on the viBe algorithm. For pixels in the current frame image that are determined to be background, a sample position is randomly selected from the corresponding pixel background sample set and replaced with the current value of the corresponding pixel. Then, a pixel position is randomly selected in the spatial neighborhood of the corresponding pixel and the current value is written into its pixel background sample set. Erosion and dilation operations are performed on the foreground mask image, and the foreground mask image and the set of dynamically connected regions are used as the modeling output of the current frame image.

4. The anti-drone observation and tracking system according to claim 1, characterized in that, The target tracking module includes: Before executing the Mean-Shift tracking algorithm, a foreground mask image is generated based on the viBe algorithm, and a pixel probability density distribution function is constructed for the pixel positions marked as foreground in the mask image. The above probability density distribution function is converted into a kernel function to initialize the weight distribution. The initial influence area of ​​the kernel function is set only within the foreground region to generate the foreground constraint initialization window. The foreground constraint initialization window is set as the initial observation window. Based on the color density kernel model corresponding to the target region in the previous frame, three-dimensional joint color space modeling is performed to generate the color histogram of the region to be matched. Construct a joint similarity kernel function to fuse the color distribution histogram with the channel feature vectors to form a dual-channel feature response; Perform gradient ascent iteration on the local region centered on the current window in the current frame image, and update the target observation window position based on the direction of the maximum value of the kernel response function until the stopping threshold condition is met; Output the updated target observation window and record the center coordinates and response function value of the corresponding window.

5. The anti-drone observation and tracking system according to claim 1, characterized in that, The consistency scoring module includes: The consistency scoring module receives the corrected target observation window coordinate information from the previous frame image and the current frame image, calculates the area of ​​each window based on the coordinates of the upper left and lower right corners of the two windows, and constructs the position overlap rate based on the area of ​​the intersection region and the area of ​​the union region. The position overlap rate is the ratio of the area of ​​the intersection region to the area of ​​the union region. The consistency scoring module receives the target density kernel model in the corresponding frame image, uses the normalized three-dimensional color joint density histogram in each model as input features, constructs a color distribution vector space, performs cosine angle calculation accordingly, and outputs the consistency score. The consistency score is compared with a preset confidence threshold. When the consistency score is less than the low confidence threshold, a failure flag is output. When the consistency score is between the low confidence threshold and the high confidence threshold, an unstable flag is output. When the consistency score is higher than the high confidence threshold, a normal flag is output, and a tracking status flag is generated.

6. The anti-drone observation and tracking system according to claim 1, characterized in that, The recapture module includes: The recapture module receives the tracking status flag. When the tracking status flag is a failure flag or an unstable flag, it determines that the target observation window in the current frame image is in an abnormal state. The recapture module reverts to the historical observation window with the highest consistency score in the most recent consecutive frames, and combines the foreground mask map of the current frame image to perform candidate region generation operations in the spatial region where the historical observation window is located and its surrounding extended region; Spatial aggregation based on connectivity clustering is performed on the region covered by the foreground mask map in the current frame image to extract multiple candidate regions, and pre-screening is performed based on the location parameters, size parameters and spatial proximity of the historical observation window of each candidate region; For all pre-screened candidate regions, calculate the corresponding joint color density histogram and perform cosine similarity calculation with the color distribution vector after normalization of the density kernel model in the historical observation window; The candidate region corresponding to the maximum value in the cosine similarity calculation is taken as the optimal target position in the current frame image, and the spatial center point of the corresponding region is taken as the updated target observation window.

Images