An unmanned aerial vehicle detection and tracking method based on motion contrast imaging
By employing a motion contrast imaging-based UAV detection and tracking method, and utilizing time-frequency analysis and multi-feature fusion, the problems of false alarms and missed detections in UAV visual detection under complex backgrounds are solved, enabling the identification and stable tracking of camouflaged or unknown targets.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHONGBEI UNIV
- Filing Date
- 2026-05-14
- Publication Date
- 2026-07-31
AI Technical Summary
Existing UAV visual inspection methods are prone to false alarms in complex dynamic backgrounds, have a high rate of missed detections under low signal-to-noise ratios, and are difficult to identify camouflaged or unknown targets. Existing motion contrast imaging technology cannot achieve stable tracking.
A motion contrast imaging-based method is adopted, which combines motion contrast imaging, image segmentation and morphological processing with time-frequency analysis and adaptive thresholding algorithm to generate motion contrast images of enhanced targets. The identification and tracking of UAV targets are achieved through support vector machine classifier and multi-feature fusion data association.
It effectively suppresses background noise, improves the detection accuracy of UAVs in complex backgrounds, can identify camouflaged or unknown targets, achieves stable tracking and identification, reduces false alarm rate, and improves detection capability in strong interference environments.
Smart Images
Figure CN122493369A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computer vision and image processing technology, specifically relating to a method for detecting and tracking unmanned aerial vehicles (UAVs) based on motion contrast imaging. Background Technology
[0002] With the widespread adoption of drone technology, the effective monitoring and identification of drones in airspace has become increasingly important. Especially in complex natural environments (such as forests and clouds) or urban settings, drones are small, have flexible trajectories, and are easily affected by dynamic background interference, posing significant challenges to traditional sensing technologies. Current mainstream drone detection technologies primarily rely on radar, radio detection, and optical imaging. Among these, visual methods based on visible light or infrared imaging have received widespread attention due to their passive, covert, and information-rich advantages. However, existing visual detection methods based on deep learning or traditional image processing rely heavily on the target's apparent features. These methods have inherent flaws. They are sensitive to dynamic backgrounds; time-varying backgrounds such as swaying leaves and drifting clouds can produce edges and textures similar to the target in the spatial domain, leading to a large number of false alarms and a sharp drop in performance. They also have weak detection capabilities for targets with low signal-to-noise ratios. When the drone is far away or there is interference such as smoke or haze, the target has an extremely low signal-to-noise ratio in the image and lacks effective appearance features for extraction, resulting in a high false negative rate. Furthermore, they have difficulty identifying camouflaged or unknown targets. The algorithms rely on the appearance of targets in the training data and have poor generalization ability for new or visually camouflaged drones that have not been seen before.
[0003] To address these issues, a novel method is urgently needed that can robustly extract moving target features from complex backgrounds without relying on the static appearance of the target. Motion contrast imaging technology offers a new approach, but existing research focuses primarily on principle verification and trajectory visualization, and a complete, engineered technical solution integrating high-sensitivity detection, accurate identification, and stable tracking for specific targets such as UAVs has not yet been developed.
[0004] Therefore, it is necessary to invent a method for detecting and tracking unmanned aerial vehicles (UAVs) based on motion contrast imaging. Summary of the Invention
[0005] To address the problems of existing UAV visual detection methods that rely on static appearance features of targets, are prone to false alarms in complex dynamic backgrounds, have high false negative rates under low signal-to-noise ratios, and are difficult to identify camouflaged or unknown targets, this invention provides a UAV detection and tracking method based on motion contrast imaging.
[0006] This invention is achieved using the following technical solution:
[0007] A method for detecting and tracking unmanned aerial vehicles (UAVs) based on motion contrast imaging includes the following steps:
[0008] S1: Perform motion contrast imaging on the input video image sequence containing the UAV target, extract the pixel grayscale time-series signal and perform time-frequency analysis, and generate a motion contrast image that enhances the target and suppresses the background by selecting at least two feature frequency bands related to the mechanical motion characteristics of the UAV for filtering and reconstruction.
[0009] S2: Perform image segmentation and morphological processing on the motion contrast image, and achieve preliminary localization of the UAV target through connected component analysis to obtain the target bounding box and centroid coordinates;
[0010] S3: Based on the target bounding box, extract the temporal signal of the corresponding region from the original image sequence, calculate its motion frequency domain feature descriptor, and input it into the classifier for recognition to obtain the target category and confidence level;
[0011] S4: Combining the target bounding box, centroid coordinates, target category, confidence level, and motion frequency domain feature descriptor, perform multi-feature fusion data association and trajectory management, and output continuous target trajectories with identity identifiers and category labels.
[0012] Furthermore, in step S1, the time-frequency analysis employs variational mode decomposition, specifically including:
[0013] S11: Construct a constrained variational model, introduce a quadratic penalty factor and augmented Lagrange multipliers to solve the variational model, and obtain a series of narrowband eigenmode functions;
[0014] S12: Perform a Hilbert transform on each of the intrinsic mode functions to obtain the analytic signal and its instantaneous frequency;
[0015] S13: Construct a time-frequency distribution map based on the analytical signal;
[0016] S14: Based on the mechanical motion characteristics of the UAV, select at least two characteristic frequency bands for bandpass filtering and image reconstruction.
[0017] Furthermore, the at least two characteristic frequency bands include:
[0018] The first frequency band, ranging from [3Hz, 20Hz], is used to extract the overall motion features of the target.
[0019] The second frequency band, ranging from [50Hz to 500Hz], is used to extract regular micro-motion features inside the target.
[0020] Furthermore, in step S2, the image segmentation adopts an adaptive thresholding algorithm, which determines the segmentation threshold by maximizing the inter-class variance, and performs morphological opening and closing operations on the binary image.
[0021] Furthermore, in step S3, the motion frequency domain feature descriptor includes at least one of the following:
[0022] The dominant frequency is the median of the instantaneous frequencies of the eigenmode function with the highest energy among all the eigenmode functions obtained from the decomposition.
[0023] Frequency stability is the standard deviation of the dominant frequency;
[0024] Motion energy contrast is the ratio of the average motion energy within the target area to the average motion energy of the local background.
[0025] Furthermore, in step S3, the classifier is a support vector machine, whose decision function is based on the feature vector constructed from the motion frequency domain feature descriptor, and outputs the target category and the corresponding confidence level.
[0026] Furthermore, in step S4, the multi-feature fusion data association and trajectory management adopts a globally optimal matching based on the Hungarian algorithm. For the i-th predicted trajectory and the j-th current frame detection result, the association cost function is defined as:
[0027] ;
[0028] in, This represents the geometric distance between the predicted bounding box and the detected bounding box. This represents the feature distance between motion frequency domain feature descriptors. Indicates classification confidence level. The normalized weighting coefficients satisfy... .
[0029] Furthermore, the typical value of the normalized weighting coefficients is... =0.4, =0.4, =0.2.
[0030] Furthermore, step S4 also includes a trajectory smoothing step: Gaussian filtering is applied to the state sequence of each confirmed trajectory within the sliding time window to eliminate jitter and false associations.
[0031] A UAV target detection and tracking system based on motion contrast imaging, comprising:
[0032] The motion contrast imaging module is used to extract pixel grayscale time-series signals from the input video image sequence and perform time-frequency analysis. By selecting at least two feature frequency bands related to the mechanical motion characteristics of the UAV for filtering and reconstruction, a motion contrast image that enhances the target and suppresses the background is generated.
[0033] The target localization module is used to perform image segmentation and morphological processing on the motion contrast image, and to obtain the bounding box and centroid coordinates of the target through connected component analysis.
[0034] The target recognition module is used to extract the temporal signal of the corresponding region from the original image sequence based on the bounding box, calculate the motion frequency domain feature descriptor, and output the target category and confidence level according to the classifier;
[0035] The multi-target tracking module is used to combine the bounding box, centroid coordinates, target category, confidence level and motion frequency domain feature descriptor to perform multi-feature fusion data association and trajectory management, and output continuous target trajectories with identity identifiers and category labels.
[0036] Compared with existing technologies, this invention provides a method for detecting and tracking unmanned aerial vehicles (UAVs) based on motion contrast imaging, which has the following advantages:
[0037] 1. This invention utilizes motion contrast imaging to actively filter out dynamic background noise in the time-frequency domain, fundamentally solving the problem of traditional methods generating numerous false alarms under complex dynamic backgrounds (such as swaying leaves or drifting clouds). By analyzing the accumulated energy of time-series signals, it can effectively enhance weak motion signals, enabling the detection of long-range, low-contrast UAVs and significantly improving the detection probability and effective range.
[0038] 2. The identification method of this invention based on motion frequency domain features (dominant frequency, frequency stability, motion energy contrast) does not rely on the static appearance features of the UAV (such as paint, model, shape), but describes the mechanical motion nature of the UAV. It can effectively distinguish UAVs from interference objects such as birds, and has the potential to identify unknown models or disguised targets.
[0039] 3. This invention integrates a multi-feature association cost function that combines geometric information, motion frequency domain features, and classification confidence, along with Kalman filter prediction and Hungarian algorithm global matching, to achieve stable identity maintenance and continuous trajectory output for multiple UAV targets. Even when targets are briefly occluded or move in overlapping directions, it can effectively reduce identity switching and trajectory breaks.
[0040] 4. This invention forms a complete technical loop for specific UAV applications, from motion contrast imaging, target localization, feature extraction and recognition to multi-target tracking. It solves the shortcomings of existing motion contrast imaging technology, which can only display trajectories but cannot perform identity recognition and continuous tracking. It achieves a continuous technical breakthrough in "seeing, recognizing accurately, and tracking steadily" highly maneuverable and weak targets such as UAVs in strong interference environments. Attached Figure Description
[0041] Figure 1 This is the overall flowchart of the present invention. Detailed Implementation
[0042] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0043] To clearly describe the algorithm flow, the following coordinate system and general parameters are first defined:
[0044] Image pixel coordinate system I: Origin (0,0) is the top left corner of the image, with the u-axis pointing to the right and the v-axis pointing downwards. Let the total number of frames in the input video sequence be T, and the shooting frame rate be... (Unit: Hz), the time corresponding to the k-th frame is The pixel coordinates in the image are (x, y), and the sequence of its gray value changes over time is denoted as g(x, y, t).
[0045] A method for UAV detection and tracking based on motion contrast imaging, such as Figure 1 As shown, it includes the following steps:
[0046] S1: Perform motion contrast imaging on the input video image sequence containing the UAV target, extract the pixel grayscale time-series signal and perform time-frequency analysis, and generate complementary motion contrast images by selecting at least two feature frequency bands with different physical meanings related to the mechanical motion characteristics of the UAV for filtering and reconstruction, so as to simultaneously enhance the overall translational motion and internal regular micro-motion characteristics of the target.
[0047] Specifically, for the input video sequence, a sliding time window of 5 frames centered on the current frame is taken, and the gray-level time-series signal g(x,y,t) of each pixel (x,y) in the window is extracted, where t changes continuously within the window. The time-frequency analysis employs Variational Mode Decomposition (VMD), a completely non-recursive adaptive signal decomposition method. It constructs and solves a constrained variational problem to decompose the original signal into a specified number of intrinsic mode functions (IMFs), which are then transformed by Hilbert transform to obtain physically meaningful instantaneous frequencies. This method is particularly suitable for processing non-stationary signals generated by UAV motion and complex backgrounds. The time-frequency analysis using VMD specifically includes:
[0048] S11: Construct a constrained variational model, introduce a quadratic penalty factor and augmented Lagrange multipliers to solve the variational model, and obtain a series of narrowband eigenmode functions. The constructed constrained variational model is as follows:
[0049] ;
[0050] The constraints are ,in Let be the Dirac distribution function. Represents the convolution operation, j is the imaginary unit, and i is the index of the intrinsic mode function. Let be the center frequency of the i-th mode.
[0051] S12: Perform a Hilbert transform on each of the intrinsic mode functions to obtain the analytic signal. ,but The amplitude is Phase is Instantaneous frequency is .
[0052] S13: Construct a time-frequency distribution map based on the analytical signal;
[0053] S14: Based on the mechanical motion characteristics of the UAV, at least two characteristic frequency bands are selected for bandpass filtering and image reconstruction. The at least two characteristic frequency bands include:
[0054] The first frequency band, ranging from [3Hz to 20Hz], is the region with the most concentrated energy in the spectrum. It covers the main motion frequencies caused by typical UAV maneuvers (such as hovering, straight flight, and hovering micro-movements) and is used to extract the overall motion characteristics of the target.
[0055] The second frequency band, ranging from 50Hz to 500Hz, represents the interval with the smallest instantaneous frequency variance and is used to extract regular micro-motion characteristics within the target. This frequency band was determined based on two considerations: the rotor speed of small UAVs is typically in the range of 3000-20000 rpm (50-333 Hz), and multiplying this by the number of blades results in rotor flicker frequencies falling within this range; statistical studies of bird flapping frequencies show that, except for hummingbirds (which can reach 50 Hz), the flapping frequencies of most birds are below 50 Hz. Therefore, the 50-500 Hz frequency band can serve as an important characteristic frequency band for distinguishing UAVs from birds.
[0056] Filtering and reconstruction are performed on these two frequency bands respectively to generate complementary motion contrast images. The reconstruction formula is:
[0057] ;
[0058] in , These are the lower / upper cutoff frequencies for the bandpass filter. The length of the image sequence used for reconstruction is such that target regions with specific motion frequencies are significantly enhanced in the generated image, while background noise is greatly suppressed.
[0059] S2: Perform image segmentation and morphological processing on the motion contrast image, and achieve preliminary localization of the UAV target through connected component analysis to obtain the target bounding box and centroid coordinates.
[0060] The image segmentation employs an adaptive thresholding algorithm, determining the optimal segmentation threshold T by maximizing the inter-class variance. The formula for calculating the inter-class variance is:
[0061] ;
[0062] in, For inter-class variance, These represent the proportions of foreground and background pixels, respectively. These are the average gray levels of the foreground and background, respectively. For the global average gray level of a motion-contrast image, the optimal threshold T is to make the gray level of the motion-contrast image equal to the global average gray level of the motion-contrast image. The maximum value. Then, by applying a threshold T, the motion contrast image is converted into a binary image:
[0063] ;
[0064] Where I(x,y) is the gray value of the motion-contrast image at pixel (x,y).
[0065] Then the binary image Morphological opening (erosion followed by dilation) and closing (dilation followed by erosion) operations are performed sequentially to eliminate boundary points and minor noise, fill small holes within the target region, and smooth the boundaries. Next, connected component analysis is performed on the processed binary image: all connected white regions in the binary image are identified, and the pixel set of each connected region is output. , where k = 1, 2, ..., K, and K is the number of detected candidate targets. For each connected region Take the minimum and maximum values of its pixel coordinates to generate the bounding box of the outer rectangle. And calculate the centroid coordinates ;
[0066] in, They are respectively regions The minimum and maximum values of the pixel coordinates. For the region The number of pixels in the image.
[0067] S3: Based on the target bounding box, extract the temporal signal of the corresponding region from the original image sequence, calculate the motion frequency domain feature descriptor that can characterize the motion essence of the UAV, and input it into a pre-trained classifier for recognition to obtain the target category and confidence level.
[0068] Specifically, for each candidate target, the corresponding gray-scale temporal signal of that region is extracted from the original video sequence based on its bounding box. The same time-frequency analysis method as S1 (variational mode decomposition and Hilbert transform) is used for analysis, and the following motion frequency domain feature descriptors are calculated:
[0069] The dominant frequency is the median instantaneous frequency of the eigenmode function with the highest energy among all the eigenmode functions obtained from the decomposition. That is:
[0070] ;
[0071] in, It is the energy of the j-th IMF. The IMF serial number with the highest energy.
[0072] Frequency stability is the standard deviation of the dominant frequency, i.e. .
[0073] Motion energy contrast is the ratio of the average motion energy within the target region to the average motion energy of the local background, i.e. ,in The average kinetic energy within the region. The average motion energy is the local background.
[0074] The above features are combined into a feature vector. The feature vectors are input into a classifier, which is a support vector machine (SVM), and its decision function is... ,in, It is the predicted category label. It is a weight vector. It is a kernel function mapping, and b is a bias term.
[0075] The classifier outputs the probability distribution P of the feature vector belonging to each preset category, with a confidence level of . The probabilities are calculated using Platt scaling: , where A and B are the fitting parameters.
[0076] S4: Combining the target bounding box, centroid coordinates, target category, confidence level, and motion frequency domain feature descriptor, inter-frame target matching and trajectory association are performed through a multi-feature data association algorithm. The initialization, confirmation, prediction, update, and termination management of the trajectory are implemented, and finally, a smooth and continuous target trajectory with a unique identifier and category label is output.
[0077] This step includes motion prediction based on Kalman filtering, data association through multi-feature fusion, trajectory update and management, temporal consistency verification, and trajectory smoothing.
[0078] Motion prediction: Maintain a Kalman filter for each activated trajectory. Define the target's state vector at time k as follows: ,in Using the centroid coordinates, For velocity. Using a linear constant velocity model, the state prediction equation is:
[0079] ;
[0080] ;
[0081] in, The state prediction value at time k. Here is the state transition matrix. Let be the state covariance matrix. Let be the process noise covariance matrix.
[0082] Multi-feature fusion data association: The Hungarian algorithm is used to perform a globally optimal match between all detections and all predicted trajectories in the current frame. For the... The predicted trajectory and the first For each detection result in the current frame, define its association cost. for:
[0083] ;
[0084] in:
[0085] The bounding box geometric similarity distance is calculated using the intersection-over-union (IoU) distance between the predicted and detected boxes: .
[0086] For motion frequency domain feature descriptor similarity, cosine distance is used: ,in The eigenvectors defined in S3 .
[0087] This represents the confidence level of the target class output by the classifier in S3.
[0088] The normalized weighting coefficients satisfy... This is used to balance the contributions of geometric features, appearance features, and classification confidence to the association, with a typical value of [value missing]. .
[0089] Find the set of matching pairs with minimum total cost using the Hungarian algorithm. Unmatched trajectories are marked as temporarily lost, and unmatched detections are marked as newly appearing candidates.
[0090] Track Update and Management:
[0091] For successfully matched trajectory-detection pairs Using the test results Update the Kalman filter (including centroid coordinates and bounding box):
[0092] ;
[0093] ;
[0094] ;
[0095] in, For the observation matrix, To observe the noise covariance matrix, The Kalman gain is used. Simultaneously, the motion frequency domain feature descriptor of the trajectory is updated to an exponentially weighted moving average of the current detected features and historical features.
[0096] For continuous Frames (e.g., frames 2-3) are all marked as newly appearing candidates, and their confidence level is higher than the threshold. Upon detection, the trajectory is initialized as a new trajectory and assigned a unique identifier. The new trajectory is initially... If the association continues to be successfully established within the frame, the status will change to confirmed.
[0097] For confirmed temporary mismatches in the trajectory, update using the predicted state and increment the temporary mismatch count. When the trajectory is temporarily lost count Exceeding the maximum allowed value or in the recent The average confidence level within the frame is below a threshold At that point, the trajectory was terminated and deleted.
[0098] Trajectory smoothing: To eliminate brief jitters and false associations, each confirmed trajectory is smoothed within a sliding time window. The state sequence within the trajectory is smoothed. A one-dimensional Gaussian filter is used to smooth the centroid coordinate sequence, and the consistency of trajectory history motion characteristics (such as velocity and dominant frequency stability) is used to verify and correct abnormal correlation points. Finally, a smooth, stable, and continuous target trajectory with identification and category labels is output.
[0099] A UAV target detection and tracking system based on motion contrast imaging, comprising:
[0100] The motion contrast imaging module is used to extract pixel grayscale time-series signals from the input video image sequence and perform time-frequency analysis. By selecting at least two feature frequency bands related to the mechanical motion characteristics of the UAV for filtering and reconstruction, a motion contrast image that enhances the target and suppresses the background is generated.
[0101] The target localization module is used to perform image segmentation and morphological processing on the motion contrast image, and to obtain the bounding box and centroid coordinates of the target through connected component analysis.
[0102] The target recognition module is used to extract the temporal signal of the corresponding region from the original image sequence based on the bounding box, calculate the motion frequency domain feature descriptor, and output the target category and confidence level according to the classifier;
[0103] The multi-target tracking module is used to combine the bounding box, centroid coordinates, target category, confidence level and motion frequency domain feature descriptor to perform multi-feature fusion data association and trajectory management, and output continuous target trajectories with identity identifiers and category labels.
[0104] In summary, this invention utilizes motion contrast imaging to actively filter out dynamic background noise in the time-frequency domain, fundamentally solving the problem of high false alarm rates in traditional methods under complex dynamic backgrounds. By analyzing the cumulative energy of time-series signals, it can effectively enhance weak motion signals, enabling the detection of UAVs at greater distances and with lower contrast compared to existing methods, thus improving the detection probability and effective range. The identification method based on motion frequency domain features is insensitive to changes in UAV paint schemes and models, and has the potential to identify unknown types or camouflaged targets.
[0105] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A method for detecting and tracking unmanned aerial vehicles (UAVs) based on motion contrast imaging, characterized in that, Includes the following steps: S1: Perform motion contrast imaging on the input video image sequence containing the UAV target, extract the pixel grayscale time-series signal and perform time-frequency analysis, and generate a motion contrast image that enhances the target and suppresses the background by selecting at least two feature frequency bands related to the mechanical motion characteristics of the UAV for filtering and reconstruction. S2: Perform image segmentation and morphological processing on the motion contrast image, and achieve preliminary localization of the UAV target through connected component analysis to obtain the target bounding box and centroid coordinates; S3: Based on the target bounding box, extract the temporal signal of the corresponding region from the original image sequence, calculate its motion frequency domain feature descriptor, and input it into the classifier for recognition to obtain the target category and confidence level; S4: Combining the target bounding box, centroid coordinates, target category, confidence level, and motion frequency domain feature descriptor, perform multi-feature fusion data association and trajectory management, and output continuous target trajectories with identity identifiers and category labels.
2. The method for detecting and tracking unmanned aerial vehicles based on motion contrast imaging according to claim 1, characterized in that, In step S1, the time-frequency analysis employs variational mode decomposition, specifically including: S11: Construct a constrained variational model, introduce a quadratic penalty factor and augmented Lagrange multipliers to solve the variational model, and obtain a series of narrowband eigenmode functions; S12: Perform a Hilbert transform on each of the intrinsic mode functions to obtain the analytic signal and its instantaneous frequency; S13: Construct a time-frequency distribution map based on the analytical signal; S14: Based on the mechanical motion characteristics of the UAV, select at least two characteristic frequency bands for bandpass filtering and image reconstruction.
3. The method for detecting and tracking unmanned aerial vehicles based on motion contrast imaging according to claim 2, characterized in that, The at least two characteristic frequency bands include: The first frequency band, ranging from [3Hz, 20Hz], is used to extract the overall motion features of the target. The second frequency band, ranging from [50Hz to 500Hz], is used to extract regular micro-motion features inside the target.
4. The method for detecting and tracking unmanned aerial vehicles based on motion contrast imaging according to claim 1, characterized in that, In step S2, the image segmentation adopts an adaptive thresholding algorithm, which determines the segmentation threshold by maximizing the inter-class variance, and performs morphological opening and closing operations on the binary image.
5. The method for detecting and tracking unmanned aerial vehicles based on motion contrast imaging according to claim 2, characterized in that, In step S3, the motion frequency domain feature descriptor includes at least one of the following: The dominant frequency is the median of the instantaneous frequencies of the eigenmode function with the highest energy among all the eigenmode functions obtained from the decomposition. Frequency stability is the standard deviation of the dominant frequency; Motion energy contrast is the ratio of the average motion energy within the target area to the average motion energy of the local background.
6. The method for detecting and tracking unmanned aerial vehicles based on motion contrast imaging according to claim 1, characterized in that, In step S3, the classifier is a support vector machine, whose decision function is based on the feature vector constructed from the motion frequency domain feature descriptor, and outputs the target category and the corresponding confidence score.
7. The method for detecting and tracking unmanned aerial vehicles based on motion contrast imaging according to claim 1, characterized in that, In step S4, the multi-feature fusion data association and trajectory management adopts a globally optimal matching based on the Hungarian algorithm. For the i-th predicted trajectory and the j-th current frame detection result, the association cost function is defined as: ； in, This represents the geometric distance between the predicted bounding box and the detected bounding box. This represents the feature distance between motion frequency domain feature descriptors. Indicates classification confidence level. The normalized weighting coefficients satisfy... .
8. The method for detecting and tracking unmanned aerial vehicles based on motion contrast imaging according to claim 7, characterized in that, The typical value of the normalized weighting coefficients is =0.4, =0.4, =0.
2.
9. The method for detecting and tracking unmanned aerial vehicles based on motion contrast imaging according to claim 1, characterized in that, Step S4 also includes a trajectory smoothing step: Gaussian filtering is applied to the state sequence of each confirmed trajectory within the sliding time window to eliminate jitter and false correlations.
10. A UAV target detection and tracking system based on motion contrast imaging, characterized in that, include: The motion contrast imaging module is used to extract pixel grayscale time-series signals from the input video image sequence and perform time-frequency analysis. By selecting at least two feature frequency bands related to the mechanical motion characteristics of the UAV for filtering and reconstruction, a motion contrast image that enhances the target and suppresses the background is generated. The target localization module is used to perform image segmentation and morphological processing on the motion contrast image, and to obtain the bounding box and centroid coordinates of the target through connected component analysis. The target recognition module is used to extract the temporal signal of the corresponding region from the original image sequence based on the bounding box, calculate the motion frequency domain feature descriptor, and output the target category and confidence level according to the classifier; The multi-target tracking module is used to combine the bounding box, centroid coordinates, target category, confidence level and motion frequency domain feature descriptor to perform multi-feature fusion data association and trajectory management, and output continuous target trajectories with identity identifiers and category labels.