Artificial Intelligence-Based Traffic Anomaly Event Recognition Method Based on UAV High-Altitude Perspective

Abstract

This invention discloses an AI-based traffic anomaly event recognition method based on a high-altitude perspective of unmanned aerial vehicles (UAVs), specifically in the field of traffic management. It addresses the challenge of identifying anomalies caused by parallax distortion in high-speed cruising scenarios by optimizing parallax distortion through continuous frame motion intensity baseline-driven flight control and zoom coordination; stabilizing target contours using a dynamic imaging window; constructing a sparse confidence kernel to analyze multi-dimensional motion; adaptively filtering low-quality candidates and refining anomaly coordinates; dynamically adjusting flight attitude and framing parameters according to the weight of anomaly-dense areas to achieve both high-altitude wide-angle coverage and fine-grained single-vehicle monitoring; eliminating parallax artifacts through confidence analysis; generating coordinates after anomaly candidate verification and driving flight control fine-tuning; and collaboratively compressing sensor bandwidth and improving alarm confidence across processing links to shorten the decision chain latency of detection, location, and tracking. This ensures coordinate accuracy and timely delivery under complex lighting and severe parallax conditions, providing reliable intelligent perception capabilities for traffic management.

Description

Technical Field

[0001] This invention relates to the field of traffic management, and more specifically, to an artificial intelligence-based method for identifying traffic anomalies based on the aerial perspective of unmanned aerial vehicles (UAVs). Background Technology

[0002] In high-speed cruise mode, the drone flies parallel to the expressway at a stable ground speed. The onboard camera collects traffic video from a top-down perspective. In order to balance panoramic coverage and the clarity of individual vehicles, the flight altitude needs to be matched with the field of view and pixel resolution. At this time, the displacement of each vehicle in the picture increases sharply between adjacent frames, the spatial features show strong parallax distortion, the motion trajectory jumps across frames, and it is far away from the existing training distribution of the convolutional network.

[0003] However, existing drone traffic anomaly identification processes lack adaptive speed and resolution adjustment strategies for high-speed cruise parallax scenarios. The motion compensation module cannot align with violently drifting targets in real time, leading to rapid inaccuracy of the detection box. The initial inference can only roughly delineate suspected anomaly clusters and cannot output precise coordinates. Alarm confidence remains consistently low, making backend scheduling decisions prone to delays or misjudgments.

[0004] To address the aforementioned problems, a technical solution is provided. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of existing technologies, embodiments of the present invention provide an AI-based traffic anomaly event recognition method based on a high-altitude perspective of an unmanned aerial vehicle (UAV). This method optimizes parallax distortion by driving flight control and zoom coordination through a continuous frame motion intensity baseline; it stabilizes the target outline using a dynamic imaging window, constructs a sparse confidence kernel to analyze multi-dimensional motion, adaptively filters low-quality candidates, and refines anomaly coordinates; flight attitude and framing parameters are dynamically adjusted according to the weight of anomaly-dense areas, achieving both high-altitude wide-angle coverage and fine-grained monitoring of individual vehicles; confidence analysis eliminates parallax artifacts, and coordinates are generated after anomaly candidate verification to drive flight control fine-tuning; and the collaborative processing links compress sensor bandwidth, improve alarm confidence, and shorten the decision chain latency of detection, location, and tracking, ensuring coordinate accuracy and push timeliness under complex lighting and severe parallax conditions. This provides reliable intelligent perception capabilities for traffic management, thus solving the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] S1: Real-time acquisition of continuous frames and recording of target centroid coordinates and initial disparity gradient histogram, output to monitoring buffer;

[0008] S2: Based on the monitoring buffer content, predict the achievable speed window and resolution level, and send speed and zoom commands to the flight controller;

[0009] S3: The flight controller performs step adjustments according to the speed window, and the zoom processor synchronously corrects the optical focal length and outputs a compensated frame stream;

[0010] S4: First, extract the five-frame sliding window disparity gradient histogram and the cross-frame vector of the target centroid from the compensation frame stream. After obtaining key features, use sparse learning to obtain disparity confidence. Based on this, filter out abnormal candidates and correct the center coordinate set.

[0011] S5: Draw an abnormally dense area map based on the corrected center coordinate set. When the weight of continuous grid points increases, adjust the flight attitude to enter the lock-follow state and report the precise coordinates in real time.

[0012] In a preferred embodiment, step S1 includes the following:

[0013] The camera equipment on the drone collects traffic videos from a top-down perspective in real time, generating a sequence of consecutive frames arranged in chronological order. For each frame in the sequence, a deep learning-based target detection algorithm is used to identify vehicle targets, generate bounding boxes, and calculate the centroid coordinates of the vehicle targets. The optical flow field between two adjacent frames is calculated using a dense optical flow algorithm, and then the magnitude of the displacement gradient is calculated to generate an initial disparity gradient histogram.

[0014] In a preferred embodiment, step S1 further includes the following:

[0015] The centroid coordinate set and initial disparity gradient histogram of each frame are recorded and output to the monitoring buffer, which adopts a first-in-first-out management method.

[0016] In a preferred embodiment, step S2 includes the following:

[0017] Based on the target centroid coordinate set and initial disparity gradient histogram in the monitoring buffer, the centroid displacement amplitude and peak change rate of the disparity gradient histogram of the vehicle target in consecutive frame images are calculated. Based on the centroid displacement amplitude and peak change rate, the achievable speed window of the UAV and the resolution level of the camera are predicted. The predicted achievable speed window and resolution level are encapsulated into instructions and sent to the flight control system to achieve adaptive adjustment of flight speed and camera optical focal length.

[0018] In a preferred embodiment, step S3 includes the following:

[0019] After receiving the achievable speed window and resolution setting instructions, the flight control system performs a step adjustment of the flight speed. First, it obtains the current flight speed. If the current flight speed is within the achievable speed window range, it maintains the current flight speed. If the current flight speed is lower than the minimum value of the achievable speed window, it adjusts the flight speed to the minimum value of the achievable speed window. If the current flight speed is higher than the maximum value of the achievable speed window, it adjusts the flight speed to the maximum value of the achievable speed window. At the same time, the zoom processor calculates the target focal length corresponding to the target resolution based on the resolution setting and adjusts the optical focal length to the target focal length.

[0020] In a preferred embodiment, step S3 further includes the following:

[0021] The adjusted flight speed and optical focal length are used to control the drone's flight and the camera to capture video frames, generating a compensated frame stream to reduce parallax distortion.

[0022] In a preferred embodiment, step S4 includes the following:

[0023] The compensation frame stream is processed using a five-frame sliding window to extract the disparity gradient histogram and the target centroid cross-frame vector from five consecutive frames. The second-order difference histogram is calculated based on the disparity gradient histogram, and the mean kurtosis is obtained to obtain the disparity amplitude index. The instantaneous energy is extracted by applying the Hilbert-Huang transform based on the target centroid cross-frame vector, and the peak value of the instantaneous energy is obtained to obtain the displacement jump factor.

[0024] In a preferred embodiment, step S4 further includes the following:

[0025] The disparity amplitude index and displacement jump factor are input into the sparse autoencoder, and the disparity confidence is output. Abnormal candidate vehicles are screened based on the disparity confidence, and the center coordinate set of abnormal candidate vehicles is corrected by calculating the average value of the centroid coordinates within five frames of the sliding window.

[0026] In a preferred embodiment, step S5 includes the following:

[0027] The traffic scene is divided into a grid based on the corrected center coordinate set. The number of abnormal targets in each grid point is counted and the grid point weight is calculated to generate an abnormal dense area map. For each grid point, the weight change rate within a continuous time window is calculated.

[0028] In a preferred embodiment, step S5 further includes the following:

[0029] When the weight change rate of adjacent grid points all exceed the preset threshold and shows an upward trend, the grid point region with the most significant increase in weight change rate is identified, the drone's flight attitude is adjusted so that the camera's viewpoint is aligned with the geometric center of the grid point region, and the drone enters a lock-and-follow state. In this state, the centroid coordinates of the abnormal target in the most recent few frames of images are calculated by weighted average to obtain the coordinates and reported in real time.

[0030] The technical effects and advantages of this invention's AI-based traffic anomaly event identification method based on a drone's high-altitude perspective are as follows:

[0031] This invention drives flight control and zoom coordination through continuous frame motion intensity baselines, achieving real-time convergence of parallax distortion in high-speed cruise scenarios. Simultaneously, it stabilizes target contours using dynamic imaging windows, constructs a sparse confidence kernel for multi-dimensional motion observation, adaptively filters low-quality candidates, and refines abnormal coordinates. Flight attitude and framing parameters are controlled and locked according to the weight of densely populated abnormal areas, achieving parallel high-altitude wide-angle coverage and fine-grained single-vehicle monitoring. Dual-exponential confidence analysis actively eliminates parallax artifacts, and abnormal candidates are verified to generate coordinate sets that drive secondary fine-tuning of flight control. Each processing link is interconnected, compressing sensor bandwidth and enhancing alarm confidence discrimination, comprehensively shortening the latency of the detection, location, and follow-up decision chain. This maintains coordinate accuracy and push timeliness for abnormal behavior on highways under complex lighting and severe parallax conditions, providing stable and reliable intelligent perception capabilities for traffic management. It also retains computational margins, allowing seamless collaboration with existing ground monitoring platforms to form an integrated air-ground operational chain. Attached Figure Description

[0032] Figure 1 This is a flowchart illustrating the AI-based traffic anomaly event identification method based on the aerial perspective of an unmanned aerial vehicle (UAV) according to the present invention. Detailed Implementation

[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] Figure 1 This invention presents an artificial intelligence-based traffic anomaly event identification method based on the aerial perspective of an unmanned aerial vehicle (UAV), comprising:

[0035] S1: Real-time acquisition of continuous frames and recording of target centroid coordinates and initial disparity gradient histogram, output to monitoring buffer.

[0036] S2: Based on the monitoring buffer content, predict the achievable speed window and resolution level, and send speed and zoom commands to the flight controller.

[0037] S3: The flight controller performs step adjustments according to the speed window, and the zoom processor synchronously corrects the optical focal length and outputs a compensated frame stream.

[0038] S4: First, extract the five-frame sliding window disparity gradient histogram and the cross-frame vector of the target centroid from the compensation frame stream. After obtaining the key features, use sparse learning to obtain the disparity confidence. Based on this, filter out abnormal candidates and correct the center coordinate set.

[0039] S5: Draw an abnormally dense area map based on the corrected center coordinate set. When the weight of continuous grid points increases, adjust the flight attitude to enter the lock-follow state and report the precise coordinates in real time.

[0040] On highways and other expressways, the rapid identification of traffic anomalies (such as accidents, congestion, or violations) is crucial for improving road safety and traffic efficiency. Traditional ground-based monitoring equipment, limited by fixed perspectives and limited coverage, struggles to detect anomalies in large-scale scenarios in a timely manner. Unmanned aerial vehicle (UAV) technology, with its high-altitude perspective and high maneuverability, can collect real-time traffic video data covering a wide range of road sections, providing a new tool for traffic monitoring. However, in high-speed cruising mode, the UAV flies at high speeds, and the video data collected by the airborne camera from above experiences a sudden increase in vehicle displacement between adjacent frames, resulting in strong parallax distortion and discontinuous motion trajectories, making traditional image processing algorithms ineffective. Therefore, a specialized adaptive identification method for UAVs' high-altitude perspective is needed to address the parallax distortion problem through real-time acquisition and analysis of video frame data, providing a reliable foundation for subsequent anomaly detection.

[0041] Step S1, as the starting point of the entire solution, is responsible for real-time acquisition of continuous frames and extraction of key features, laying the data foundation for subsequent flight control adjustments and anomaly identification. In high-speed cruise scenarios, the processing logic of step S1 must ensure the accuracy and real-time performance of the data to address the challenges posed by parallax distortion.

[0042] The goal of step S1 is to acquire continuous frame images of traffic video in real time using the camera equipment on the UAV, extract the target centroid coordinates and initial disparity gradient histogram of the vehicles, and output them to the monitoring buffer. This process provides basic data for subsequent steps (such as S2 predicting the speed window and resolution level), ensuring that the UAV can dynamically adjust its flight parameters according to the real-time traffic scene. The specific technical logic of step S1 is as follows.

[0043] Step S1 includes the following:

[0044] S1.1, Real-time acquisition of continuous frame images:

[0045] The drone flies parallel to the expressway in high-speed cruise mode, its onboard camera continuously capturing overhead traffic video at fixed time intervals, generating a sequential sequence of images arranged chronologically. Each frame is sequentially labeled as frame one, frame two, and so on, up to the current frame. Each frame has a fixed width and height, recording visual information about the road and vehicles. The camera operates at a capture frequency of 30 times per second to ensure the continuous frame sequence reflects the dynamic changes in the traffic scene. The purpose of real-time acquisition of the continuous frame sequence is to provide dynamic visual data for subsequent vehicle target detection and parallax analysis, ensuring the system can accurately capture the real-time changing characteristics of the traffic scene.

[0046] S1.2, Vehicle target detection and centroid coordinate calculation:

[0047] For each frame in a continuous image sequence, a deep learning-based object detection algorithm is used to identify vehicle targets. Specifically, a pre-trained neural network model is used to determine the location of the vehicle targets by analyzing image pixel features. The detection result is a bounding box for each vehicle, defined by the coordinates of its top-left and bottom-right corners, representing the area of ​​the vehicle target. Based on the bounding box coordinates of each vehicle, the centroid coordinates are calculated. This is done by adding the horizontal coordinates of the top-left and bottom-right corners of the bounding box and averaging them to obtain the horizontal centroid position, and then adding the vertical coordinates and averaging them to obtain the vertical centroid position. The centroid coordinates represent the center position of the vehicle target in the current frame image and are used for subsequent analysis of the vehicle target's motion in the continuous image sequence.

[0048] The deep learning-based target detection algorithm can accurately identify vehicle targets in each frame of the image, providing high-precision bounding box information and ensuring reliable quantification of vehicle positions. Calculating centroid coordinates simplifies the representation of vehicle target positions, facilitating subsequent steps to track vehicle trajectory and supporting the identification of traffic anomalies.

[0049] S1.3, Calculation of the initial disparity gradient histogram:

[0050] To quantify the disparity change between adjacent frames in a continuous image sequence, an initial disparity gradient histogram from the previous frame to the current frame is calculated. First, the pixel motion of the previous and current frames is analyzed using a dense optical flow algorithm to generate an optical flow field. This field records the horizontal and vertical displacement components of each pixel between the two frames. Next, for each pixel in the optical flow field, the magnitude of its displacement gradient is calculated. The squares of the horizontal and vertical displacement components are added together, and the square root of the sum is extracted to obtain the pixel's displacement gradient value, representing the pixel's motion intensity in the continuous image sequence. Then, the displacement gradient values ​​of all pixels are statistically analyzed to generate a histogram. The horizontal axis of the histogram represents the range of displacement gradient values, divided into multiple continuous intervals, while the vertical axis represents the number of pixels falling into each interval. The initial disparity gradient histogram describes the disparity distribution characteristics of the current frame image pair and is used to analyze the disparity change trend caused by the high-speed flight of the UAV.

[0051] By calculating the optical flow field using a dense optical flow algorithm, the motion information of pixels in a continuous sequence of image frames can be comprehensively captured, reflecting the parallax changes caused by the high-speed flight of the UAV. The generation of the initial parallax gradient histogram can statistically quantify the parallax distribution characteristics between image frames, providing a scientific basis for adjusting the UAV's flight speed and the camera's resolution, and improving the stability of the system processing.

[0052] S1.4, Data output to the monitoring buffer:

[0053] The processing results of each frame in a continuous sequence of images, including the centroid coordinates of all detected vehicle targets and the initial disparity gradient histogram, are recorded and output to the monitoring buffer. The monitoring buffer is a temporary storage area with a fixed capacity, employing a first-in, first-out (FIFO) data management method. When the monitoring buffer is full, the oldest stored frame image data is removed, and the processing results of the new frame image are added. The monitoring buffer stores the frame image itself, the corresponding set of vehicle target centroid coordinates, and the initial disparity gradient histogram, ensuring that subsequent processing steps can access the processing results of the most recent frames, supporting dynamic analysis of traffic scenes.

[0054] Step S1 provides dynamic visual data and feature information for the entire traffic anomaly recognition system by acquiring a continuous frame image sequence in real time and performing vehicle target detection, disparity analysis, and data storage. The acquisition of the continuous frame image sequence provides the raw input for subsequent processing, vehicle target detection and centroid coordinate calculation provide positional information, the calculation of the initial disparity gradient histogram provides disparity change characteristics, and the storage in the monitoring buffer integrates this information and passes it to subsequent steps.

[0055] Step S1 has completed the real-time acquisition of continuous frame images, recorded the target centroid coordinates and initial disparity gradient histogram, and output these data to the monitoring buffer, laying the data foundation for subsequent processing. However, in high-speed cruise mode, when the UAV flies along the expressway, the inter-frame displacement of vehicles in the image increases sharply, leading to aggravated disparity distortion. If the flight speed and camera resolution cannot be dynamically adjusted according to the traffic scene, it will affect the target detection accuracy. Step S2 aims to predict the achievable speed window and resolution level of the UAV by analyzing the data in the monitoring buffer, and issue commands to the flight control system to achieve adaptive control of flight parameters, providing stable image input for subsequent steps.

[0056] Step S2 takes the target centroid coordinates and initial disparity gradient histogram in the monitoring buffer as input to predict the UAV's achievable speed window and camera resolution level, and sends speed and zoom commands to the flight control system. The goal is to reduce the impact of disparity distortion on target detection in high-speed cruise mode by dynamically adjusting flight speed and resolution, ensuring that subsequent steps S3 to S5 can accurately identify traffic anomalies based on stable image data.

[0057] Step S2 includes the following:

[0058] S2.1, Extract monitoring buffer data:

[0059] First, data is extracted from the monitoring buffer. The monitoring buffer stores the processing results of the most recent frames, including the target centroid coordinate set and the initial disparity gradient histogram for each frame. The target centroid coordinate set is a list of coordinates containing the centroid position of each vehicle in the image, used for subsequent analysis of vehicle displacement in consecutive frames. The initial disparity gradient histogram is a distribution map of pixel displacement gradients in each frame, used to analyze the disparity variation trend between frames. The extraction process is completed by reading the latest record from the monitoring buffer, ensuring that the data reflects the real-time traffic scene status.

[0060] S2.2, Calculate the displacement amplitude of the target centroid:

[0061] Next, the target centroid displacement amplitude is calculated to quantify the average displacement of the vehicle across consecutive frames. The specific calculation process is as follows: For two consecutive frames, firstly, for each vehicle, the centroid coordinates in the current frame and the previous frame are obtained. Then, the difference between the two sets of coordinates is calculated to obtain the horizontal and vertical displacement components. Next, for each vehicle, the magnitude of its displacement vector is calculated. This is done by squared the horizontal and vertical displacement components, summing the sum, and then taking the square root of the sum to obtain the vehicle's displacement magnitude. Finally, the average displacement magnitude of all vehicles is calculated by summing the displacement magnitudes of all vehicles and dividing by the total number of vehicles to obtain the target centroid displacement amplitude.

[0062] The calculation of the target centroid displacement amplitude reflects the intensity of the vehicle's motion in consecutive frames of images. This metric quantifies the impact of vehicle displacement on parallax distortion during high-speed drone flight. By accurately measuring the average displacement, it is possible to scientifically determine whether the current flight speed is appropriate, thus providing an objective basis for adjusting the drone's flight speed and ensuring the quality of image data.

[0063] S2.3, Analyze the peak value changes of the disparity gradient histogram:

[0064] Further analysis of the peak value variation in the initial disparity gradient histogram is conducted to assess the dynamic characteristics of the disparity distribution. The specific process is as follows: For each frame's initial disparity gradient histogram, its peak value is extracted, which is the displacement gradient value with the highest pixel count in the histogram. Next, the rate of change of the peak value between two consecutive frames is calculated. This is done by taking the absolute value of the difference between the current frame's peak value and the previous frame's peak value, and then dividing this difference by the previous frame's peak value. The peak value change rate represents the magnitude of the disparity distribution's change over time and is used to determine the stability or worsening trend of disparity distortion.

[0065] Analyzing the peak variation of the initial disparity gradient histogram can capture the dynamic characteristics of disparity distribution. The rate of change of the peak directly reflects the severity of disparity distortion, allowing for an assessment of the difficulty in capturing image details based on real-time data. This analysis provides a basis for adjusting camera resolution, ensuring the clarity and usability of image data in different traffic scenarios.

[0066] S2.4, Predicted reachable speed window:

[0067] Based on the target centroid displacement amplitude and the peak rate of change of the initial disparity gradient histogram, the achievable speed window of the UAV is predicted, i.e., the suggested minimum and maximum speeds. The specific prediction logic is as follows: First, the target centroid displacement amplitude and peak rate of change are comprehensively evaluated. When the target centroid displacement amplitude is large and the peak rate of change is high, it indicates that the vehicle displacement is too large and the disparity distortion is severe, in which case the flight speed needs to be reduced. When the target centroid displacement amplitude is small and the peak rate of change is low, it indicates that the displacement and disparity changes are stable, in which case the flight speed can be maintained or increased. The calculation method is as follows: the minimum speed is obtained by multiplying the current flight speed by a reduction factor, the magnitude of which is proportional to the product of the target centroid displacement amplitude and the peak rate of change; the maximum speed is obtained by multiplying the current flight speed by an increase factor, the magnitude of which is proportional to the complement of the target centroid displacement amplitude and the peak rate of change.

[0068] The complement is defined as a fixed value minus the product of the target centroid displacement amplitude and the peak rate of change.

[0069] The predicted reachable speed window enables the adjustment of the UAV's flight speed in real time based on changes in the traffic scene. This process accurately determines the severity of parallax distortion by combining the target centroid displacement amplitude and peak rate of change, thereby determining a reasonable speed range. The adjusted flight speed effectively reduces the impact of parallax distortion on target detection and ensures the stability of image data.

[0070] S2.5, Predicted resolution level:

[0071] Based on the peak change rate of the initial disparity gradient histogram, the camera's resolution level is predicted. The specific prediction logic is as follows: a high peak change rate indicates drastic disparity changes, requiring a higher resolution to capture more image details; a low peak change rate indicates stable disparity changes, allowing for a lower resolution to optimize computational efficiency. The calculation method is as follows: the adjusted resolution level is obtained by multiplying the current resolution by an adjustment factor, the magnitude of which is proportional to the peak change rate. This method determines the appropriate resolution level for the current scene.

[0072] The predicted resolution level dynamically adjusts the camera resolution based on the drastic changes in parallax. Increasing the resolution when parallax changes drastically ensures accurate capture of image details; decreasing the resolution when parallax changes are stable optimizes the utilization efficiency of computing resources. This mechanism improves adaptability in different scenarios and ensures a balance between image data quality and processing efficiency.

[0073] S2.6, issue commands to the flight control system:

[0074] Finally, the predicted achievable speed window and resolution setting are encapsulated into commands and sent to the flight control system. Specifically, the minimum speed, maximum speed, and resolution setting data are integrated into a single control command package, which is transmitted to the flight control system via the communication interface. Upon receiving the commands, the flight control system adjusts the UAV's flight speed based on the minimum and maximum speeds and adjusts the camera's optical focal length based on the resolution setting, thereby achieving adaptive control of flight parameters.

[0075] By issuing commands to the flight control system, real-time adjustments to flight speed and camera resolution were achieved. This function ensures that the drone can stably acquire video data according to changes in traffic scenarios during high-speed cruise mode. The adjusted parameters effectively reduce the impact of parallax distortion, providing high-quality image input for subsequent traffic anomaly event identification.

[0076] Step S2 extracts the target centroid coordinate set and initial disparity gradient histogram from the monitoring buffer, calculates the target centroid displacement amplitude and peak rate of change, and then predicts the achievable speed window of the UAV and the camera resolution level. Finally, it sends commands to the flight control system. This achieves adaptive control of flight parameters in high-speed cruise mode, reduces the impact of sudden increases in vehicle inter-frame displacement and disparity distortion on image quality, and provides stable image data support for subsequent steps.

[0077] Step S2 has completed the analysis of the monitoring buffer content. Based on the target centroid coordinates of consecutive frames and the initial parallax gradient histogram, the achievable speed window of the UAV and the resolution level of the camera are predicted, and speed and zoom commands are issued to the flight control system. However, in high-speed cruise mode, when the UAV flies parallel to the expressway and the onboard camera collects traffic video from above, the target displacement between adjacent frames increases sharply due to the need to match the flight altitude and field of view with the pixel resolution, resulting in significant parallax distortion. To ensure the accuracy of subsequent abnormal event identification, step S3 needs to adjust the flight speed and optical focal length in real time according to the instructions in step S2 to generate a stable compensated frame stream, thereby reducing the impact of parallax distortion on target detection and providing reliable image data for step S4.

[0078] The specific processing logic of step S3 is based on the achievable speed window and resolution level output in step S2. Through the coordinated work of the flight control system and zoom processor, the flight speed of the UAV and the optical focal length of the camera are adjusted to generate a compensated frame stream with a stable target outline. This reduces the parallax distortion caused by the sudden increase in target displacement in high-speed cruise mode and ensures the accuracy of subsequent traffic anomaly event identification.

[0079] Step S3 includes the following:

[0080] S3.1, receive achievable speed window and resolution level commands:

[0081] The flight control system first receives the command issued after analyzing the monitoring buffer content in step S2. The command includes two parts: the achievable speed window and the resolution setting. The achievable speed window refers to the range of suggested flight speeds for the UAV predicted based on the target centroid coordinates and the initial parallax gradient histogram in consecutive frames. The resolution setting refers to the camera resolution setting standard corresponding to the target resolution. The flight control system stores the achievable speed window as the basis for adjusting the flight speed, and simultaneously transmits the resolution setting to the zoom processor as the basis for adjusting the optical focal length.

[0082] S3.2, Perform step adjustment of flight speed:

[0083] The drone's flight speed is adjusted based on the received reachable speed window. The adjustment process involves the following steps: First, the drone's current flight speed is acquired and compared to the reachable speed window. If the current flight speed is within the reachable speed window, it remains unchanged without any adjustment. If the current flight speed is below the minimum value of the reachable speed window, the flight speed is increased to the minimum value. If the current flight speed is above the maximum value of the reachable speed window, the flight speed is decreased to the maximum value. The adjusted flight speed is recorded by the flight control system and used to control the drone's actual flight status.

[0084] By comparing the current flight speed with the range of the achievable speed window and making stepwise adjustments, the drone's flight speed is ensured to always remain within the predicted recommended range. This adjustment method effectively reduces the sudden increase in target displacement and parallax distortion caused by excessively high or low flight speeds, thereby improving the stability of acquired video frames and providing reliable image data for subsequent anomaly event identification.

[0085] S3.3, Synchronous correction of optical focal length:

[0086] The zoom processor adjusts the camera's optical focal length based on the received resolution level. The specific adjustment process is as follows: First, the zoom processor determines the standard value corresponding to the target resolution based on the resolution level. Then, the zoom processor acquires the camera's current optical focal length and compares it with the standard focal length corresponding to the target resolution. Based on the comparison result, the zoom processor calculates the difference between the current optical focal length and the target focal length and gradually adjusts the current optical focal length to the target focal length. The adjusted optical focal length is recorded by the zoom processor and used to control the camera's zoom settings.

[0087] The optical focal length is adjusted according to the resolution setting to ensure that the camera captures video frames at a focal length that matches the target resolution. This synchronous correction method ensures that the image resolution is adapted to the flight altitude and field of view, improving the sharpness of the target outline and image quality, thereby enhancing the accuracy of subsequent traffic anomaly identification.

[0088] S3.4, Generate the compensated frame stream:

[0089] After adjusting the flight speed and zoom processor to the adjusted optical focal length, the camera continuously acquires traffic video frames using the adjusted flight speed and optical focal length as parameters, generating a compensated frame stream. The compensated frame stream is a sequence of video frames that, after adjusting the flight speed and optical focal length, reduces parallax distortion and stabilizes the target outline. The generated compensated frame stream is stored and directly transmitted to step S4 as data input for subsequent processing.

[0090] By generating a compensated frame stream using adjusted flight speed and optical focal length, parallax distortion caused by a sudden increase in target displacement during high-speed cruise mode can be effectively reduced, while maintaining the stability of the target outline. This compensated frame stream provides high-quality image data for step S4 to extract the five-frame sliding window parallax gradient histogram and the target centroid cross-frame vector, ensuring the coordinate accuracy and detection reliability of anomaly event recognition.

[0091] Step S3 receives the achievable speed window and resolution setting instructions from step S2. The flight control system adjusts the flight speed based on the comparison between the current flight speed and the achievable speed window. The zoom processor calculates and adjusts the optical focal length based on the resolution setting, and finally generates a compensated frame stream with the adjusted parameters. This process effectively solves the interference of parallax distortion on target detection in high-speed cruise mode, provides stable image data for step S4, and ensures the overall performance and reliability of the traffic anomaly recognition system.

[0092] Step S3, through the coordinated scheduling of the flight control system and the zoom processor, adjusted the UAV's flight speed and the camera's optical focal length based on the speed window and resolution level predicted in step S2, generating a compensated frame stream. This compensated frame stream effectively reduces parallax distortion caused by the mismatch between the field of view and pixel resolution in high-speed cruise mode. However, in complex traffic scenarios and under varying lighting conditions, the compensated frame stream may still contain residual motion trajectories of abnormal targets jumping across frames or parallax artifacts, affecting the accurate identification and localization of subsequent abnormal events. To address this issue, step S4 uses the compensated frame stream as input, employs feature extraction and sparse learning to calculate the parallax confidence score, filter out abnormal candidates, and correct the center coordinate set, providing reliable abnormal target coordinate support for step S5.

[0093] Step S4 receives the compensation frame stream output from step S3, aiming to generate a corrected center coordinate set by extracting key features and performing anomaly identification. Based on the five-frame sliding window disparity gradient histogram and the target centroid cross-frame vector, the disparity amplitude exponent and displacement jump factor are calculated. A sparse autoencoder is used to generate disparity confidence, and finally, anomaly candidates are filtered and the coordinates of the anomalous target are accurately located, providing a data foundation for the anomaly dense area map drawing and flight attitude adjustment in step S5.

[0094] Step S4 includes the following:

[0095] S4.1, Extract the five-frame sliding window disparity gradient histogram:

[0096] The input compensation frame stream undergoes a five-frame sliding window process. Specifically, the current frame, its two preceding frames, and its two following frames are selected from the compensation frame stream, totaling five frames. For each of these five frames, a disparity gradient histogram is calculated between each pair of adjacent frames. The disparity gradient histogram is generated by first analyzing the displacement gradient distribution of each pixel in the adjacent frames, and then statistically analyzing the frequency of these displacement gradients to form a histogram. The displacement gradient distribution refers to the amount of displacement change of a pixel in the horizontal and vertical directions. Using this method, four pairs of adjacent frames are generated from the five frames, corresponding to four sets of disparity gradient histograms, each set representing the disparity statistical characteristics of a pair of adjacent frames.

[0097] The reason for using a five-frame sliding window is that five consecutive images can completely capture the dynamic changes in vehicle movement within a traffic scene, avoiding the omission of key information in single-frame or two-frame analysis. The disparity gradient histogram accurately reflects the characteristics of inter-frame disparity by statistically analyzing the distribution of displacement gradients, providing reliable data support for subsequent analysis of disparity fluctuations.

[0098] S4.2, Extract the target centroid cross-frame vector:

[0099] Based on a five-frame sliding window, the centroid coordinates of each vehicle in each frame are extracted. Centroid coordinates refer to the horizontal and vertical positions of the vehicle's geometric center in the image. Next, the difference in centroid coordinates between adjacent frames is calculated to generate a target centroid cross-frame vector. The calculation process involves subtracting the centroid coordinates of the previous frame from the centroid coordinates of the next frame to obtain the horizontal and vertical displacement components. For four pairs of adjacent frames within the five-frame sliding window, four sets of target centroid cross-frame vectors are generated, each representing the displacement of a vehicle between adjacent frames.

[0100] The purpose of extracting the target centroid cross-frame vector is to quantify the motion trajectory of each vehicle in consecutive frames of images. By calculating the difference in centroid coordinates, the direction and magnitude of vehicle displacement can be accurately described. This provides direct evidence for subsequent analysis of vehicle motion anomalies, such as detecting sudden acceleration or deceleration. The generation of the target centroid cross-frame vector ensures a comprehensive understanding of vehicle dynamics and improves the sensitivity of anomaly event identification.

[0101] S4.3, Calculate the parallax amplitude index:

[0102] Based on the aforementioned four sets of disparity gradient histograms, a second-order difference histogram is calculated. Specifically, the disparity gradient histograms corresponding to three adjacent frames are selected, and these three sets of histograms are differentially processed one by one. First, the difference between the first and second sets of histograms is calculated, then the difference between the second and third sets is calculated, and so on, ultimately generating three sets of second-order difference histograms. Next, the kurtosis is calculated for each set of second-order difference histograms. The kurtosis is calculated by first finding the average of all statistical values ​​in the histogram, then calculating the fourth square of each statistical value relative to the average, then calculating the second square of each statistical value relative to the average, and finally taking the ratio of the fourth square to the second square as the kurtosis. The disparity amplitude exponent is defined as the sum of the kurtosis values ​​of the three sets of second-order difference histograms, followed by the average.

[0103] The purpose of calculating the second-order difference histogram is to capture the dynamic trend of the disparity gradient histogram over time, while the introduction of kurtosis quantifies the sharpness and fluctuation characteristics of the disparity distribution. The disparity amplitude index, by averaging three sets of kurtosis values, comprehensively reflects the overall magnitude of disparity change within a five-frame sliding window.

[0104] S4.4, Calculate the displacement jump factor:

[0105] For each vehicle, a vector time series is constructed using the four sets of target centroid cross-frame vectors within a five-frame sliding window. Then, the Hilbert-Huang transform is applied to decompose this vector time series and extract instantaneous energy. The Hilbert-Huang transform process involves first decomposing the vector time series into several intrinsic mode components, then performing a Hilbert transform on each component, calculating the squared amplitude at each time point, and obtaining the instantaneous energy time series. The displacement jump factor is defined as the maximum value found in the instantaneous energy time series, i.e., the peak value of the vehicle's instantaneous energy within the five-frame sliding window.

[0106] The Hilbert-Huang transform is suitable for analyzing non-stationary and nonlinear signals, and can effectively decompose the dynamic characteristics of vehicle motion. The extraction of instantaneous energy reflects the intensity change of vehicle displacement, while the selection of peak values ​​directly quantifies the maximum degree of motion abrupt changes. This method scientifically identifies abnormal vehicle motion behaviors, such as sudden stops or sharp turns, ensuring the system can accurately detect abrupt events in traffic scenarios and improving the accuracy of anomaly identification.

[0107] S4.5, using a sparse autoencoder to obtain disparity confidence:

[0108] The disparity amplitude index and displacement jump factor of each vehicle are combined into a feature vector, which serves as the input data. This feature vector is then processed by a pre-trained sparse autoencoder. The sparse autoencoder encodes and decodes the input feature vector using a multi-layer neural network, forcibly introducing sparsity constraints during encoding to extract a sparse representation of the feature vector. The final output is a value between 0 and 1, called the disparity confidence score. The disparity confidence score represents the degree of confidence in the disparity of a vehicle within the current five-frame sliding window.

[0109] The sparse autoencoder extracts key features from the disparity amplitude exponent and displacement abruptness factor through unsupervised learning to generate sparse representations, avoiding interference from redundant information. The output of disparity confidence provides a quantitative standard for subsequent screening of anomaly candidates. It can adaptively identify anomalous patterns in disparity and motion features, improving the ability to distinguish anomalous events and ensuring the reliability of subsequent processing.

[0110] S4.6, Filter out anomalous candidates and correct the center coordinate set:

[0111] Based on the disparity confidence of each vehicle, a predefined threshold is set for filtering. If a vehicle's disparity confidence is lower than this threshold, it is marked as an anomalous candidate vehicle. For vehicles marked as anomalous candidates, their centroid coordinate set is corrected. Specifically, the correction method is to calculate the average of the five centroid coordinates of the vehicle within a five-frame sliding window, that is, to add the horizontal coordinates of the five frames and take the average, and add the vertical coordinates of the five frames and take the average, to obtain a new coordinate value, which is used as the corrected centroid coordinate.

[0112] By filtering out abnormal candidate vehicles using disparity confidence, erroneous targets caused by disparity artifacts or motion anomalies can be effectively eliminated. The average of five centroid coordinates is used when correcting the center coordinate set, reducing the impact of noise or errors in single-frame images on the localization results.

[0113] Step S4 uses the compensated frame stream output from step S3 as input data. By extracting the five-frame sliding window disparity gradient histogram and the target centroid cross-frame vector, it calculates the disparity amplitude exponent and displacement jump factor. Then, it uses a sparse autoencoder to generate disparity confidence, ultimately completing the filtering of anomaly candidates and the correction of the center coordinate set. The corrected center coordinate set is used as input for step S5 to draw anomaly dense area maps and adjust the UAV's flight attitude, realizing a complete transition from image analysis to practical application.

[0114] Step S4 successfully filtered out anomaly candidates and output the corrected center coordinate set by performing feature extraction and sparse learning on the compensated frame stream. This coordinate set accurately locates abnormal targets in the traffic scene, laying the foundation for subsequent anomaly event identification and response. However, in high-speed cruise mode, relying solely on the corrected center coordinate set is insufficient to meet the dynamic traffic anomaly event management needs under complex lighting and severe parallax conditions. Therefore, step S5 uses the corrected center coordinate set as input, draws anomaly dense area maps, analyzes the grid point weight change trends, and adjusts the UAV's flight attitude in real time to lock onto and follow abnormal targets, while simultaneously reporting precise coordinates to provide efficient and accurate decision support for traffic management departments.

[0115] Step S5 takes the corrected center coordinate set output from step S4 as input and aims to adjust the UAV's flight attitude to enter a lock-and-follow state by constructing an anomaly dense area map and monitoring the dynamic changes in grid point weights, and to extract and report the precise coordinates of the anomaly target in real time. Its core task is to achieve dynamic identification and continuous tracking of high-incidence areas of anomalies in high-speed cruise scenarios, ensuring that the UAV can stably monitor anomalies and support rapid response from traffic management departments.

[0116] Step S5 includes the following:

[0117] S5.1, Draw an abnormally dense area map:

[0118] The traffic scene is divided into multiple uniform grid points, each covering a fixed area. Next, the number of anomalous targets within each grid point is counted, and this number is divided by the grid area to calculate the weight value of each grid point. Finally, the weight values ​​of all grid points are combined to form an anomaly density map, reflecting the spatial distribution of anomalous targets.

[0119] By dividing the number of anomalous targets by the grid area to calculate the weight value, the spatial distribution density of anomalous targets can be quantified, thus intuitively showing the concentration of anomalous events in different areas. The grid-based mesh generation and weight calculation method is simple and efficient, quickly identifying areas with high incidence of anomalous events and providing clear regional basis for subsequent UAV flight attitude adjustments and coordinate reporting.

[0120] S5.2, Monitor the weight changes of continuous grid points:

[0121] For each grid point, the changes in its weight value are recorded within consecutive time windows. Specifically, the difference between the grid point weight value in the current time window and the grid point weight value in historical time windows is calculated. Then, this difference is divided by the number of frames contained in the time window to obtain the grid point weight change rate, which is used to characterize the dynamic change trend of abnormal target density.

[0122] By calculating the grid weight change rate, the trend of abnormal target density changing over time can be captured, allowing for the timely detection of areas with significantly increased density. This dynamic monitoring method enhances the system's sensitivity to the evolution of abnormal events, enabling UAVs to quickly perceive the aggregation or diffusion of abnormal targets in traffic scenarios. Compared to static analysis, the introduction of the weight change rate provides stronger support for real-time detection of abnormal events, ensuring that the system can take appropriate measures in the early stages of anomalies.

[0123] S5.3, Adjust flight attitude to enter lock-on follow mode:

[0124] When the weight change rate of adjacent grid points is detected to exceed a preset threshold and show an upward trend, the grid point region with the most significant increase in weight change rate is further identified. Subsequently, the flight attitude of the UAV is adjusted so that its onboard camera's viewpoint is aligned with the geometric center of this region, and the relative position between the UAV and the abnormal target is kept stable by controlling the position and angle of the UAV, thereby entering a lock-and-follow state.

[0125] By identifying the grid region with the highest weight change rate and adjusting the flight attitude to lock onto that region, the UAV can continuously track the high-incidence locations of abnormal events. This avoids monitoring blind spots caused by improper flight attitude and improves the system's ability to stably monitor abnormal targets in complex traffic scenarios. Simultaneously, maintaining the stability of the relative position with abnormal targets ensures greater continuity and reliability in subsequent coordinate calculations and reporting.

[0126] S5.4, real-time reporting of precise coordinates:

[0127] When the drone is in lock-and-follow mode, the centroid coordinates of the abnormal target in the most recent few frames are processed. The processing method is as follows: different weight values ​​are assigned to each frame based on the time distance, with the weight value of the frame image that is closer to the current time being larger; then, a weighted average of these centroid coordinates is calculated to obtain the precise coordinates of the abnormal target, and these coordinates are reported in real time.

[0128] By calculating a weighted average of the centroid coordinates of multiple image frames and determining the weight values ​​inversely proportional to the time distance, this method effectively smooths the coordinate data and reduces errors introduced by noise, illumination variations, or parallax distortion in a single image frame, thereby improving the accuracy and stability of the reported coordinates. Especially in high-speed drone cruising mode, it significantly reduces the impact of external environmental interference on target positioning, ensuring high reliability of the reported coordinate data and providing accurate spatial information support for anomaly management in traffic scenarios.

[0129] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0130] It should be noted that the system of the present invention can be deployed on the device itself to realize embedded applications, or it can run on a PC or other terminal with a user interface, thereby meeting a variety of hardware environments and usage requirements.

[0131] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0132] It should be noted that, in this document, the use of relational terms such as "first" and "second" is merely to distinguish one entity or operation from another, and does not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0133] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for identifying traffic anomalies based on the aerial perspective of unmanned aerial vehicles (UAVs), characterized in that, Including the following steps: S1: Real-time acquisition of continuous frame image sequences. For each frame in the continuous frame image sequence, a deep learning-based target detection algorithm is used to identify vehicle targets, generate bounding boxes, and calculate the centroid coordinates of the vehicle targets, which are the center positions of the vehicle targets in the current frame image. The optical flow field of adjacent two frames is calculated using a dense optical flow algorithm, and then the magnitude of the displacement gradient is calculated to generate an initial disparity gradient histogram, which is a statistical distribution of the magnitude of the displacement gradient. The set of centroid coordinates and the initial disparity gradient histogram of each frame image are recorded and output to the monitoring buffer. S2: Based on the target centroid coordinate set in the monitoring buffer and the initial disparity gradient histogram, calculate the centroid displacement amplitude of the vehicle target in consecutive frame images. The centroid displacement amplitude is the average displacement magnitude of the difference between the centroid coordinates of two consecutive frames. Calculate the peak change rate of the disparity gradient histogram. The peak change rate is the absolute difference between the peak values ​​of two consecutive frames divided by the peak value of the previous frame. Based on the centroid displacement amplitude and peak change rate, predict the achievable speed window of the UAV and the resolution level of the camera. The achievable speed window is calculated by multiplying the current speed by a factor related to the displacement amplitude and peak change rate. The resolution level is calculated by multiplying the current resolution by an adjustment factor proportional to the peak change rate. Encapsulate the predicted achievable speed window and resolution level into commands and send them to the flight control system. S3: After receiving the achievable speed window and resolution setting instructions, the flight control system performs a step adjustment of the flight speed, adjusting the current flight speed to within the achievable speed window range; at the same time, the zoom processor calculates the target focal length corresponding to the target resolution based on the resolution setting, and adjusts the optical focal length to the target focal length; The adjusted flight speed and optical focal length are used to control the drone's flight and the camera to capture video frames, generating a compensated frame stream. S4: The compensation frame stream is processed by a five-frame sliding window to extract the disparity gradient histogram and the target centroid cross-frame vector from five consecutive frames; the second-order difference histogram is calculated based on the disparity gradient histogram, and the kurtosis mean is obtained to obtain the disparity amplitude index. The instantaneous energy is extracted by applying Hilbert-Huang transform to the cross-frame vector of the target centroid, and the peak value of the instantaneous energy is obtained to obtain the displacement jump factor. The disparity amplitude index and the displacement jump factor are input into the sparse autoencoder, and the disparity confidence is output. The disparity confidence value between 0 and 1 represents the degree of disparity confidence. Abnormal candidate vehicles are screened based on disparity confidence. Vehicles with disparity confidence below a threshold are marked as abnormal candidates. The center coordinate set of abnormal candidate vehicles is then corrected by calculating the average centroid coordinates within five frames of the sliding window. S5: Based on the corrected center coordinate set, the traffic scene is divided into a grid of points. The number of abnormal targets in each grid point is counted and the grid point weight is calculated. The grid point weight is the number of abnormal targets divided by the grid point area, generating an abnormal dense area map. For each grid point, its weight change rate within a continuous time window is calculated. The change rate is the weight difference divided by the number of frames. When the weight change rate of adjacent grid points all exceed the preset threshold and shows an upward trend, the grid point area with the most significant increase in weight change rate is identified. The drone's flight attitude is adjusted so that the camera's view is aligned with the geometric center of the grid point area, and the lock-follow state is entered. In this state, the centroid coordinates of the abnormal targets in the most recent few frames are calculated using a weighted average. The weight of the weighted average is based on the inverse ratio of time distance to obtain accurate coordinates and report them in real time.

2. The method for identifying traffic anomalies based on an unmanned aerial vehicle (UAV) high-altitude perspective according to claim 1, characterized in that, Step S1 includes the following: The camera equipment on the drone collects traffic videos from a top-down perspective in real time, generating a sequence of consecutive frames arranged in chronological order. For each frame in the sequence, a deep learning-based target detection algorithm is used to identify vehicle targets, generate bounding boxes, and calculate the centroid coordinates of the vehicle targets. The optical flow field between two adjacent frames is calculated using a dense optical flow algorithm, and then the magnitude of the displacement gradient is calculated to generate an initial disparity gradient histogram.

3. The method for identifying traffic anomalies based on an unmanned aerial vehicle (UAV) high-altitude perspective according to claim 2, characterized in that, Step S1 also includes the following: The centroid coordinate set and initial disparity gradient histogram of each frame are recorded and output to the monitoring buffer, which adopts a first-in-first-out management method.

4. The method for identifying traffic anomalies based on an unmanned aerial vehicle (UAV) high-altitude perspective according to claim 3, characterized in that, Step S2 includes the following: Based on the target centroid coordinate set and initial disparity gradient histogram in the monitoring buffer, the centroid displacement amplitude and peak change rate of the disparity gradient histogram of the vehicle target in consecutive frame images are calculated. Based on the centroid displacement amplitude and peak change rate, the achievable speed window of the UAV and the resolution level of the camera are predicted. The predicted achievable speed window and resolution level are encapsulated into instructions and sent to the flight control system to achieve adaptive adjustment of flight speed and camera optical focal length.

5. The method for identifying traffic anomalies based on an unmanned aerial vehicle (UAV) high-altitude perspective according to claim 4, characterized in that, Step S3 includes the following: After receiving the achievable speed window and resolution setting commands, the flight control system performs a step adjustment of the flight speed. First, it obtains the current flight speed. If the current flight speed is within the achievable speed window range, it maintains the current flight speed unchanged. If the current flight speed is lower than the minimum value of the achievable speed window, it adjusts the flight speed to the minimum value of the achievable speed window. If the current flight speed is higher than the maximum value of the reachable speed window, the flight speed is adjusted to the maximum value of the reachable speed window; at the same time, the zoom processor calculates the target focal length corresponding to the target resolution based on the resolution level, and adjusts the optical focal length to the target focal length.

6. The method for identifying traffic anomalies based on an unmanned aerial vehicle (UAV) high-altitude perspective according to claim 5, characterized in that, Step S3 also includes the following: The adjusted flight speed and optical focal length are used to control the drone's flight and the camera to capture video frames, generating a compensated frame stream to reduce parallax distortion.

7. The method for identifying traffic anomalies based on an unmanned aerial vehicle (UAV) high-altitude perspective according to claim 6, characterized in that, Step S4 includes the following: The compensation frame stream is processed using a five-frame sliding window to extract the disparity gradient histogram and the target centroid cross-frame vector from five consecutive frames; the second-order difference histogram is calculated based on the disparity gradient histogram, and the kurtosis mean is obtained to get the disparity amplitude index. The instantaneous energy is extracted by applying the Hilbert-Huang transform to the cross-frame vector of the target centroid, and the peak value of the instantaneous energy is obtained to obtain the displacement jump factor.

8. The method for identifying traffic anomalies based on an unmanned aerial vehicle (UAV) high-altitude perspective according to claim 7, characterized in that, Step S4 also includes the following: The disparity amplitude index and displacement jump factor are input into the sparse autoencoder, and the disparity confidence is output. Abnormal candidate vehicles are screened based on the disparity confidence, and the center coordinate set of abnormal candidate vehicles is corrected by calculating the average value of the centroid coordinates within five frames of the sliding window.

9. The method for identifying traffic anomalies based on an unmanned aerial vehicle (UAV) high-altitude perspective according to claim 8, characterized in that, Step S5 includes the following: The traffic scene is divided into a grid based on the corrected center coordinate set. The number of abnormal targets in each grid point is counted and the grid point weight is calculated to generate an abnormal dense area map. For each grid point, the weight change rate within a continuous time window is calculated.

10. The method for identifying traffic anomalies based on an unmanned aerial vehicle (UAV) high-altitude perspective according to claim 9, characterized in that, Step S5 also Includes the following: When the weight change rate of adjacent grid points all exceed the preset threshold and shows an upward trend, the grid point region with the most significant increase in weight change rate is identified, the drone's flight attitude is adjusted so that the camera's viewpoint is aligned with the geometric center of the grid point region, and the drone enters a lock-and-follow state. In this state, the centroid coordinates of the abnormal target in the most recent few frames of images are calculated by weighted average to obtain the coordinates and reported in real time.

Images