Unmanned aerial vehicle video-based traffic signal timing parameter inversion method and device

Abstract

The application discloses a traffic signal timing parameter inversion method and equipment based on unmanned aerial vehicle video, and utilizes the traffic flow video shot by the unmanned aerial vehicle to automatically identify the release start time of each phase in combination with vehicle detection and track information, and corrects noise and abnormal values through mathematical statistics method to realize stable estimation on the signal period and the green light time of each stage. The application can automatically, efficiently and accurately restore the signal timing characteristics of the urban intersection without the traditional detector or manual observation, and provides reliable data support for traffic signal optimization and intelligent control.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation and video detection technology, and to a method and device for inverting traffic signal timing parameters based on UAV video. Background Technology

[0002] With the rapid development of Intelligent Transportation Systems (ITS), the optimization and evaluation of traffic signal control has become an important research direction for improving road traffic efficiency and traffic safety. Reasonable timing of urban road traffic signals can effectively reduce delays, lower energy consumption, and improve traffic capacity. Therefore, obtaining accurate signal control parameters is crucial for traffic engineering analysis and intelligent signal control.

[0003] Traditional signal control research primarily relies on ground-based detection equipment, such as induction coils, microwave radar, geomagnetic detectors, or video detectors, to identify vehicle traffic conditions and calculate signal cycle information. Some studies also utilize log files output by signal controllers to record the start and end times of green lights for each phase, thereby acquiring signal timing data. However, these methods have significant limitations in practical applications: firstly, the installation and maintenance costs of ground-based detection equipment are high, and construction can disrupt normal traffic flow; secondly, different brands or models of signal controllers lack a unified communication protocol and data interface, making it difficult to standardize the extraction of signal log data. Furthermore, in areas without signal controller interfaces or with missing data, the above methods cannot effectively obtain periodic timing information.

[0004] In recent years, drones have been widely used for traffic flow monitoring and data collection due to their advantages such as wide field of view, high maneuverability, and flexible deployment. Drone-captured videos can directly cover the entire intersection, allowing for the extraction of vehicle trajectories, speeds, and conflict behaviors. However, existing drone video analysis methods primarily focus on target detection and trajectory tracking, lacking effective methods for automatically identifying traffic signal control phases from videos. Since traffic light status cannot be directly observed in drone videos, traditional methods typically rely on manual annotation or external signal data to infer signal cycles, which has significant limitations in large-scale intersection monitoring and subsequent data analysis.

[0005] Furthermore, the inference during the signal phase is affected by various factors: the vehicle release areas of different phases overlap spatially; some phases, such as left-turn or mixed-traffic phases, have sparse vehicles, leading to unstable detection timing; there are transition delays between adjacent phases, such as during all-red or yellow light periods; and factors such as video frame rate and detection confidence may also introduce errors. Therefore, to achieve high-precision mapping from UAV video data to signal control parameters, a robust multi-region detection and time-series statistical inference algorithm needs to be designed. Summary of the Invention

[0006] To address the aforementioned problems, this invention proposes a method and device for inverting traffic signal timing parameters based on UAV video. This invention utilizes traffic flow video captured by UAVs, combined with vehicle detection and trajectory information, to automatically identify the start time of the green light for each phase. Furthermore, it corrects noise and outliers using mathematical statistics methods, achieving stable estimations of the signal cycle and green light times for each phase. This invention can operate independently of the signal controller, enabling intelligent perception that "directly recovers signal timing from video."

[0007] The technical solution adopted in this invention is as follows:

[0008] A method for inverting traffic signal timing parameters based on UAV video includes the following steps:

[0009] S1. Capture video of the intersection using a drone, and preprocess it to obtain a continuous video frame sequence;

[0010] S2. Divide the continuous frame video sequence into several phase detection regions in the coordinate system, use target detection and multi-target tracking algorithms to extract vehicle bounding boxes and center points, and construct vehicle temporal trajectory sequences;

[0011] S3. Based on the vehicle time-series trajectory sequence, the observation time axis is divided into several periods; within each period, for each phase detection area, the time when the vehicle first enters the phase detection area is calculated as the start time for the release of the phase detection area, and a delay constraint is applied to the temporal relationship of the phase detection area to correct the abnormal detection.

[0012] S4. Construct the boundary time series of the signal stage based on the start time of the release in each phase detection area, and calculate the green light duration, red light duration and total cycle duration for each stage;

[0013] S5. Outlier removal and robust mean estimation are performed sequentially on the stage duration datasets obtained in multiple periods, and confidence intervals are calculated through resampling to obtain the recommended timing parameters and confidence ranges for each stage;

[0014] S6. Generate and output a report containing recommended timing parameters, phase timing relationships, and abnormal cycle diagnostic information.

[0015] Furthermore, the drone uses hovering aerial photography to obtain video of the intersection, with a flight altitude of 80-120 meters and a lens pitch angle of 89°-90°.

[0016] Furthermore, the preprocessing includes the following steps:

[0017] The captured video is standardized in terms of frame rate and resolution. If there are multiple shots or video segments, the video segments are stitched together to a unified coordinate system using spatial feature matching or ground control points with the same name. The stitched video sequence is then processed for lens distortion correction, image stabilization and shaking removal, and brightness equalization, outputting a continuous video frame sequence with consistent geometry and brightness.

[0018] Furthermore, the step of extracting vehicle bounding boxes and center points using object detection and multi-object tracking algorithms, and constructing a temporal trajectory dataset for the vehicle, specifically includes:

[0019] For each frame of video, a target detector is used to obtain the vehicle bounding box, and the vehicle center point is calculated based on the vehicle bounding box. A multi-target tracking algorithm is used to perform association matching on the same vehicle in different frames based on the vehicle center point to obtain the temporal trajectory sequence of the vehicle.

[0020] Furthermore, the multi-target tracking algorithm is the ByteTrack multi-target tracking algorithm, which establishes a matching cost function between vehicles in different frames by combining Euclidean distance and intersection-over-union ratio to perform cross-frame association. The weighted cost function is:

[0021] ,

[0022] in, The matching cost between vehicle i in the current frame and vehicle j in the next frame. and These represent the coordinates of the vehicle's center point in different frames. These are the weighting coefficients. This indicates intersection, union, and ratio.

[0023] Furthermore, the method for calculating the release start time of the phase detection area is as follows:

[0024] ,

[0025] in, This indicates the start time of release for the i-th phase detection region in the k-th cycle. This represents the time interval corresponding to the k-th period. This indicates that the vehicle has entered the i-th phase detection area;

[0026] The delay constraint is:

[0027] ,

[0028] in, This indicates the start time of release for the (i+1)th phase detection region in the kth cycle. This represents the minimum conversion delay between different phase detection regions.

[0029] Furthermore, step S4 specifically includes:

[0030] The release start times of each phase detection region within each cycle are sorted to obtain the boundary time series of the signal stage;

[0031] The green light duration for each phase is calculated based on the difference between the start times of the clearance in two adjacent signal phases.

[0032] The total duration of each cycle is calculated based on the start time of the first release of two adjacent cycles;

[0033] The red light duration for each stage is obtained by subtracting the total cycle duration from the green light duration for each stage.

[0034] Furthermore, within each cycle, for each phase detection area, if no vehicle enters the phase detection area, the start time of the passage is recorded as a null value; when calculating the green light duration, if any boundary of each stage is a null value, the green light duration is recorded as a missing value.

[0035] Furthermore, step S5 includes:

[0036] For each signal phase, duration data of the signal phase is collected from multiple cycles, and missing values ​​are removed to obtain a phase duration dataset; outlier removal is performed on the phase duration dataset using the interquartile range method; based on the data samples after outlier removal, the truncated mean is calculated as the recommended timing parameter for the duration of the signal phase, and the confidence interval for the duration of the signal phase is calculated using the Bootstrap resampling method.

[0037] A traffic signal timing parameter inversion device based on UAV video, comprising:

[0038] One or more processors;

[0039] Memory, used to store one or more programs;

[0040] When the one or more programs are executed by the one or more processors, the one or more processors implement the above-described method for inverting traffic signal timing parameters based on UAV video.

[0041] This invention aims to solve the technical challenge of existing UAV traffic detection technologies being unable to automatically and accurately identify the start and end times of intersection signal phases. It proposes an algorithm model that can recover the complete signal control timing sequence solely from video trajectory data, even without external signal controller data. This method can automatically estimate the green light duration, red light duration, and cycle length for each phase under UAV visual observation conditions, ensuring high recognition accuracy, strong environmental adaptability, and robust and reliable statistical results. It possesses strong versatility and engineering application value.

[0042] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0043] 1. This invention combines wide-area visual data captured by UAVs hovering at high altitudes with deep learning detection algorithms, achieving for the first time automated inversion of traffic signal timing without the need for ground detectors. By performing high-resolution processing and multi-stage video stabilization stitching on the UAV video, it effectively solves the problems of limited viewing angle, insufficient detection range, and data discontinuity in traditional ground detection methods, achieving global observation of the entire intersection functional area and providing reliable visual data support for signal timing estimation.

[0044] 2. This invention employs a multi-target detection and tracking mechanism, enabling high-precision vehicle trajectory recognition and continuous tracking in complex traffic environments. Through dynamic trajectory clustering and release direction recognition algorithms, the system can automatically distinguish vehicle release groups corresponding to different signal phases and deduce the signal phase switching time from trajectory density change characteristics. This achieves automatic estimation of signal cycle and green light duration, significantly improving the intelligence and accuracy of signal timing recognition.

[0045] 3. This invention constructs a parameter optimization model for green light duration and cycle duration based on robust statistical estimation methods. By introducing IQR outlier removal, Bootstrap resampling, and truncated mean estimation mechanisms, the influence of short-term traffic disturbances and detection noise on the results is effectively suppressed. This robust estimation framework exhibits higher stability and robustness under complex traffic conditions compared to traditional averaging and least squares methods, achieving a green light duration error controlled within 1.5 seconds and a signal cycle estimation accuracy of 99.2%.

[0046] 4. This invention constructs a complete traffic signal timing estimation system through an integrated process of "UAV visual data acquisition—depth detection and tracking—signal phase identification—robust parameter estimation." This method can achieve rapid and automated recovery of signal timing at urban intersections without relying on traditional inductive detectors or manual data collection. It has significant advantages such as flexible deployment, strong adaptability, and high versatility, providing a novel technical approach for intelligent traffic signal control and urban traffic management. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the overall method in an embodiment of the present invention.

[0048] Figure 2 This is an aerial view of an intersection taken by a drone in an embodiment of the present invention.

[0049] Figure 3 Based on the embodiments of the present invention Figure 2 A diagram illustrating vehicle trajectory tracking. Detailed Implementation

[0050] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments, but the embodiments of the present invention are not limited to the following.

[0051] like Figure 1 As shown, a traffic signal timing parameter inversion method based on UAV video includes the following steps: S1. Capture video of the intersection using a UAV and perform preprocessing to obtain a continuous video frame sequence.

[0052] In this embodiment, the drone used is a DJI Air 2S model, and the flight mode is to hover and observe directly above the intersection, with a flight altitude of... Approximately 100 meters, camera tilt angle The camera should be positioned at approximately 89°-90°, maintaining a near-vertical overhead view to ensure full coverage of the intersection. Figure 2 As shown. The video was collected during the morning and evening rush hours on the main urban roads on June 16-17, 2025, and includes 12 independent sampling periods, with a total duration of approximately 480 minutes.

[0053] After synchronizing the original captured video with timestamps, the following processing is performed:

[0054] (1) Unified frame rate and resolution: The video frame rate is standardized to 29.97 fps and the resolution is unified to 3840×2160.

[0055] (2) Video stitching and field of view unification: If there are multiple shots or multiple video segments, a multi-view spatial feature matching algorithm based on scale-invariant feature transformation and random sampling consistency algorithm RANSAC is adopted. The algorithm extracts key feature points between video frames or multiple camera videos, calculates their corresponding matching relationships, uses RANSAC to remove outliers and estimate the homography matrix, and transforms each segment to a unified pixel coordinate system, thereby forming a seamlessly stitched panoramic overhead video sequence. After stitching, the video is processed for temporal synchronization. The timestamp differential alignment method is used to correct the frame time of different video sources to ensure the synchronization consistency of cross-view data in the time dimension.

[0056] (3) Image stabilization and geometric correction: Image stabilization algorithm is applied to remove shake and compensate for wind disturbance and slight displacement; and inverse distortion correction is applied to lens distortion to ensure that pixel coordinates have geometric consistency.

[0057] (4) Brightness / contrast equalization: Perform global or local histogram equalization on the frame sequence to improve detection robustness.

[0058] (5) Region clipping and ROI definition: Delineate the detection area based on road channelization information and remove invalid background areas.

[0059] The continuous video frame sequence after the above preprocessing and corresponding timestamp As input for subsequent detection and analysis.

[0060] S2. Divide the continuous frame video sequence into several phase detection regions in the coordinate system, use target detection and multi-target tracking algorithms to extract vehicle bounding boxes and center points, and construct vehicle temporal trajectory sequences;

[0061] In this embodiment, eight rectangular detection regions are divided under a unified pixel coordinate system. ( For each phase detection region Define boundaries:

[0062] in, Indicates the first Each phase detection region This indicates the pixel coordinates of the vehicle's center point within the video frame. Indicates the first The minimum horizontal coordinate of each detection region in the pixel coordinate system Indicates the first The maximum horizontal coordinate of each detection region in the pixel coordinate system Indicates the first The minimum vertical coordinate of each detection region in the pixel coordinate system Indicates the first The maximum vertical coordinate of each detection region in the pixel coordinate system.

[0063] To facilitate the determination of whether a vehicle has entered the phase detection area, an entry discrimination function is defined:

[0064]

[0065] when When =1, the vehicle is considered to have entered the phase detection area. In this embodiment, the target detector employs a YOLOv11s-based target detection model to identify traffic targets in UAV video frames. This model is pre-trained on the VisDrone2019 dataset and optimized through transfer learning to improve its performance in detecting small-scale vehicle targets. The detection results output include the target category, confidence score, and pixel-level bounding box coordinates.

[0066] like Figure 3 As shown, this invention extracts vehicle positions from UAV overhead images and performs trajectory correlation between consecutive frames to obtain complete motion trajectory data of vehicles at an intersection. A target detector is applied to each frame to obtain vehicle bounding boxes. And calculate the vehicle's center point:

[0067]

[0068] in, Indicates the number of detected cells in the current frame. The vehicle's serial number, This represents the time index corresponding to the video frame. Indicates vehicle At any moment The horizontal coordinates of the top-left pixel of the bounding box. Indicates vehicle At any moment The vertical coordinate of the top-left pixel of the bounding box. Indicates vehicle At any moment The width of the bounding box, Indicates vehicle At any moment The height of the bounding box, Indicates vehicle At any moment The coordinates of the center point.

[0069] To achieve continuous vehicle trajectory tracking, the ByteTrack multi-target tracking algorithm is employed. This algorithm establishes a matching cost function between detection boxes by combining Euclidean distance and intersection-over-union (IoU):

[0070]

[0071] in, Indicates the vehicle in the current frame With the next frame vehicle The matching cost, The weighting coefficient (0.3 in this embodiment) is used to balance spatial distance and similarity of overlapping areas, thereby achieving high-precision trajectory association. The algorithm combines a Kalman filter to predict the target position and uses the Hungarian matching algorithm to complete the optimal association.

[0072] The final vehicle trajectory sequence is obtained:

[0073]

[0074] in, Indicates the number is The movement trajectory of the vehicle in the video.

[0075] S3. Based on the vehicle time-series trajectory sequence, the observation time axis is divided into several periods; within each period, for each phase detection area, the time when the vehicle first enters the phase detection area is calculated as the start time for the release of the phase detection area, and a delay constraint is applied to the temporal relationship of the phase detection area to correct the anomaly detection.

[0076] Specifically, based on the extracted vehicle trajectory data, the video timeline is divided into periods and the initial release time of the phase is calculated. First, based on the observed time series and the preset signal period length... The global data is segmented, and the period length can be set to any value within the range of 120 to 160 seconds, depending on the actual intersection control strategy. If the data acquisition start time is... Then the first Period is a time interval Then, within each cycle, the detection areas of each phase are statistically analyzed. The initial release time for vehicles:

[0077]

[0078] Will Defined as the first The start time for the release of the i-th phase detection area in the cycle is if no vehicle enters during that cycle. ,but .

[0079] To avoid abnormal timing offsets caused by detection delays or tracking interruptions, a phase delay constraint is further introduced:

[0080]

[0081] in, This represents the minimum transition delay between different phase detection regions (typically 1.5 seconds). This constraint is used to correct for outlier detection samples and maintain phase order consistency.

[0082] S4. Construct the boundary time series of the signal stage based on the start time of the release in each phase detection area, and calculate the green light duration, red light duration and total cycle duration for each stage.

[0083] Specifically, for the sequence of start times of each phase Perform stage boundary and duration calculations. First, determine the release start time for each phase detection region within each cycle. Sort the data and construct the stage boundary time series. Based on this, the duration of the green light and red light for each stage is defined:

[0084]

[0085] in, For the first Mid-cycle phase Green light time This indicates the start time of the release in the (n+1)th stage of the k-th cycle. This represents the start time of the release in the nth stage of the kth cycle. If any boundary is... Then corresponding Set as a missing value.

[0086] The period duration is defined as:

[0087] ,

[0088] in, and These represent the start times of the first release in the k-th and k+1-th cycles, respectively.

[0089] Red light duration:

[0090]

[0091] in, For the first Mid-cycle phase Red light duration.

[0092] By analyzing vehicle throughput, queue length change rate, and headway trends, the location of stage switching points can be further refined to ensure the physical rationality of timing boundaries. Meanwhile, to reduce noise impact, missing boundaries should be marked and treated differently in subsequent statistics.

[0093] S5. For the stage duration datasets obtained in multiple periods, outlier removal and robust mean estimation are performed sequentially, and confidence intervals are calculated using a resampling method to obtain the recommended timing parameters and confidence ranges for each stage.

[0094] Specifically, for multi-period sample sets Conduct statistical analysis and collect data at each stage. Cross-period samples (After removing missing terms), a robust estimation process is used to obtain reliable signal duration values.

[0095] 1. (1) Tukey IQR outlier removal: calculated based on interquartile range and Preserve samples:

[0096]

[0097] like Then rollback to use .

[0098] (2) Truncated mean (recommended value): using the pruning ratio (Example) ), calculate the truncated mean

[0099] .

[0100] (3) Bootstrap confidence interval: based on Using the mother sample, perform The number of resampling iterations is typically 500-2000, preferably 1000. For each iteration, a truncated mean is calculated, and this mean is taken. and quantiles as Confidence interval:

[0101]

[0102] The final output is the recommended green light duration for each stage. and its confidence interval; and simultaneously perform the same robust estimation on the set of cycle durations to obtain the recommended cycle. This invention selects the truncated mean as the recommended timing parameter for the period and the green light duration at each stage, ensuring that abnormal data does not affect the overall trend.

[0103] S6. Generate and output a report containing recommended timing parameters, phase timing relationships, and abnormal cycle diagnostic information. Specifically, this includes:

[0104] (1) Recommended green light duration, red light duration and total cycle duration for each stage;

[0105] (2) A detailed table of the duration of the original and removed identifiers for each period;

[0106] (3) Phase timing diagram and stage switching sequence diagram;

[0107] (4) Anomaly alarm report, used to identify periods containing missing data, abnormal boundaries or noise interference.

[0108] To verify the accuracy of this invention, the inversion results were compared with those recorded by the ground signal controller. The results show that the cycle estimation accuracy reaches 99.2%, and the error in the green light duration at each stage is mostly less than 1.5 seconds, verifying the high accuracy and stability of the method of this invention.

[0109] This invention enables automatic estimation of signal timing schemes using only UAV video data, without the need for ground detectors or manual recording. This method has the following advantages:

[0110] 1. Flexible data collection with wide coverage;

[0111] 2. The algorithm is robust and suitable for various weather and traffic conditions;

[0112] 3. High estimation accuracy and controllable error;

[0113] 4. It can provide real-time input for traffic signal optimization and intelligent control.

[0114] Therefore, this invention has broad application potential in urban traffic signal assessment, timing optimization, and intelligent intersection construction.

[0115] This invention is applicable to traffic intersections without signal detectors, ground induction coils, or signal network interfaces. It utilizes a high-definition camera mounted on a drone to acquire overhead video sequences and reconstructs signal timing information for each release phase using non-contact visual observation. This enables reverse inference, periodic consistency verification, and operational status assessment of the traffic signal control system. Furthermore, this invention achieves adaptive reconstruction of signal cycles, green light times, all-red times, and phase sequences through temporal clustering and regional correlation analysis of vehicle traffic events. It further supports parallel computing scenarios across multiple intersections and features adaptive parameter updates to adapt to varying traffic flow densities, weather conditions, and aerial camera altitudes. Simultaneously, this invention maintains stable recognition accuracy under different video resolutions, lighting conditions, and camera pitch angles, enabling signal control information extraction and timing optimization from multi-source drone video under a unified spatiotemporal reference, providing highly reliable input for coordinated regional traffic signal control.

[0116] The method of this invention can be deployed on edge computing devices, traffic control center servers, or cloud computing platforms, and is implemented in Python. The system includes a video input module, a data parsing module, a detection and tracking module, a signal phase recognition module, a statistical analysis module, and a visualization output module. The system has a visual user interface that can display the detection area, vehicle trajectory, and phase recognition results in real time. It supports batch processing of UAV videos and automated report generation, and has modular expansion interfaces for linkage with intelligent signal control systems.

[0117] Those skilled in the art will understand that embodiments of the present invention can be provided as an algorithm-based software implementation, an embedded chip-based hardware implementation, or a hybrid software and hardware implementation. For example, in a software implementation, the above steps can be executed on a computing device with a processor (such as a UAV ground station or a traffic analysis server) via computer program instructions; in a hardware implementation, the algorithm of the present invention can be embedded in an edge computing module or an intelligent traffic detection terminal to achieve real-time signal phase identification and timing estimation.

[0118] This invention is described with reference to the algorithm flow and module structure according to embodiments of the invention, but is not limited to specific implementations. Details such as the order of steps, module division, and parameter thresholds can be appropriately adjusted by those skilled in the art based on different road scenarios, UAV parameters, or video resolutions. For example, under different frame rate conditions, the period segment length, delay correction parameters, and sliding window length in the signal phase recognition algorithm can all be flexibly set, and these remain within the scope of this invention.

[0119] It should be noted that in this paper, terms such as "based on UAV video," "detection area," "trajectory sequence," and "robust statistical estimation" are used for illustrative purposes only and their meanings should be interpreted broadly. For example, "UAV video" is not limited to visible light aerial video, but also includes various forms such as infrared imaging, panoramic stitching, oblique photography, and multispectral images; "trajectory sequence" can be obtained by visual inspection, radar ranging, lidar, or fusion perception systems, and its source does not affect the realization of the technical concept of this invention.

[0120] Furthermore, this invention is applicable not only to typical cross intersections but also to T-junctions and other complex nodes with signal control. For atypical signal control structures, the same timing inference function can be achieved simply by adjusting the number of area divisions and phase logic constraints. This invention can also be extended to multi-intersection collaborative analysis scenarios in urban road networks, enabling synchronous signal inference across multiple nodes through parallel computing.

[0121] Furthermore, this invention possesses excellent versatility and can be applied to multiple sub-fields of intelligent transportation, such as traffic flow analysis, signal timing assessment, traffic conflict identification, and congestion evolution modeling. In different applications, Bootstrap resampling, IQR elimination, and truncated mean estimation can be used as modular tools to form a standardized traffic data analysis workflow, further improving the accuracy and repeatability of traffic parameter estimation.

[0122] The above description is merely a preferred embodiment of the present invention. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the scope of protection of the present invention. All equivalent transformations or substitutions made based on the technical concept of the present invention should be covered within the scope of protection of the present invention. Various modifications and improvements to the form and details of the present invention made by those skilled in the art without departing from the spirit and substance of the present invention should be considered equivalent solutions of the present invention.

Claims

1. A method for inverting traffic signal timing parameters based on UAV video, characterized in that, Includes the following steps: S1. Capture video of the intersection using a drone, and preprocess it to obtain a continuous video frame sequence; S2. Divide the continuous frame video sequence into several phase detection regions in the coordinate system, use target detection and multi-target tracking algorithms to extract vehicle bounding boxes and center points, and construct vehicle temporal trajectory sequences; S3. Based on the vehicle time-series trajectory sequence, the observation time axis is divided into several periods; within each period, for each phase detection area, the time when the vehicle first enters the phase detection area is calculated as the start time for the release of the phase detection area, and a delay constraint is applied to the temporal relationship of the phase detection area to correct the abnormal detection. S4. Construct the boundary time series of the signal stage based on the start time of the release in each phase detection area, and calculate the green light duration, red light duration and total cycle duration for each stage; S5. Outlier removal and robust mean estimation are performed sequentially on the stage duration datasets obtained in multiple periods, and confidence intervals are calculated through resampling to obtain the recommended timing parameters and confidence ranges for each stage; S6. Generate and output a report containing recommended timing parameters, phase timing relationships, and abnormal cycle diagnostic information.

2. The method for inverting traffic signal timing parameters based on UAV video according to claim 1, characterized in that, The drone uses hovering aerial photography to capture video of the intersection, flying at an altitude of 80-120 meters and with a camera tilt angle of 89°-90°.

3. The method for inverting traffic signal timing parameters based on UAV video according to claim 2, characterized in that, The preprocessing includes the following steps: The captured video is standardized in terms of frame rate and resolution. If there are multiple shots or video segments, the video segments are stitched together to a unified coordinate system using spatial feature matching or ground control points with the same name. The stitched video sequence is then processed for lens distortion correction, image stabilization and shaking removal, and brightness equalization, outputting a continuous video frame sequence with consistent geometry and brightness.

4. The method for inverting traffic signal timing parameters based on UAV video according to claim 1, characterized in that, The process of extracting vehicle bounding boxes and center points using object detection and multi-object tracking algorithms, and constructing a temporal trajectory dataset for the vehicle, specifically includes: For each frame of video, a target detector is used to obtain the vehicle bounding box, and the vehicle center point is calculated based on the vehicle bounding box. A multi-target tracking algorithm is used to perform association matching on the same vehicle in different frames based on the vehicle center point to obtain the temporal trajectory sequence of the vehicle.

5. The method for inverting traffic signal timing parameters based on UAV video according to claim 4, characterized in that, The multi-target tracking algorithm described is the ByteTrack multi-target tracking algorithm. It establishes a matching cost function between vehicles in different frames by combining Euclidean distance and intersection-over-union ratio (IoU) for cross-frame association. The weighted cost function is: ， in, The matching cost between vehicle i in the current frame and vehicle j in the next frame. and These represent the coordinates of the vehicle's center point in different frames. These are the weighting coefficients. This indicates intersection, union, and ratio.

6. The method for inverting traffic signal timing parameters based on UAV video according to claim 1, characterized in that, The method for calculating the release start time of the phase detection area is as follows: ， in, This indicates the start time of release for the i-th phase detection region in the k-th cycle. This represents the time interval corresponding to the k-th period. This indicates that the vehicle has entered the i-th phase detection area; The delay constraint is: ， in, This indicates the start time of release for the (i+1)th phase detection region in the kth cycle. This represents the minimum conversion delay between different phase detection regions.

7. The method for inverting traffic signal timing parameters based on UAV video according to claim 1, characterized in that, Step S4 specifically includes: The release start times of each phase detection region within each cycle are sorted to obtain the boundary time series of the signal stage; The green light duration for each phase is calculated based on the difference between the start times of the clearance in two adjacent signal phases. The total duration of each cycle is calculated based on the start time of the first release of two adjacent cycles; The red light duration for each stage is obtained by subtracting the total cycle duration from the green light duration for each stage.

8. The method for inverting traffic signal timing parameters based on UAV video according to claim 7, characterized in that, Within each cycle, for each phase detection area, if no vehicle enters the phase detection area, the start time of the passage is recorded as a null value; when calculating the green light duration, if any boundary of each stage is a null value, the green light duration is recorded as a missing value.

9. The method for inverting traffic signal timing parameters based on UAV video according to claim 8, characterized in that, Step S5 includes: For each signal phase, duration data of the signal phase is collected from multiple cycles, and missing values ​​are removed to obtain a phase duration dataset. The IQR interquartile range method is used to remove outliers from the phase duration dataset. Based on the data samples after removing outliers, the truncated mean is calculated as the recommended timing parameter for the duration of the signal phase, and the Bootstrap resampling method is used to calculate the confidence interval of the duration of the signal phase.

10. A traffic signal timing parameter inversion device based on UAV video, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the traffic signal timing parameter inversion method based on UAV video as described in any one of claims 1 to 9.

Images