A wet road risk assessment method based on unmanned aerial vehicle identification of vehicle trajectory

Abstract

The present application belongs to the field of risk prediction, and relates to a wet road risk assessment method based on unmanned aerial vehicle identification of vehicle trajectory. The method adopts a gimbal-carrying way of high-resolution unmanned aerial vehicle, extracts video data of vehicle driving trajectory of dangerous road section, maps to obtain vehicle ground trajectory sequence, calculates parameter indexes of normal sample and wet road according to reference trajectory curve, compares wet road with normal sample to output distribution difference; based on meteorological coefficient, the threshold index on wet road is adaptively adjusted, and the weighted threshold proportion and the duration exceeding the 95% quantile of normal weather are calculated; the adjacency consistency and the spatial coherence within the section are judged, and finally the section-level risk score and risk level are formed. The accuracy, real-time performance and interpretability of risk early warning are significantly improved, and the false alarm rate caused by environmental misjudgment or equipment error is effectively reduced.

Description

Technical Field

[0001] This invention belongs to the field of risk prediction technology, specifically relating to a method for assessing the risk of slippery roads based on vehicle trajectory recognition by unmanned aerial vehicles. Background Technology

[0002] my country's transportation industry has developed rapidly, and its transportation infrastructure has been continuously improved. However, significant weaknesses remain in traffic hazard warning systems, especially on icy and snowy roads, and traffic safety on dangerous road sections in icy and snowy weather continues to face severe challenges. Investigations and analyses have revealed that in severe weather conditions where ice and snow are the primary influencing factors, the presence of snow or ice on roads creates particularly serious safety hazards for vehicles traveling on dangerous sections. Current road detection technologies and hazard warning methods cannot accurately analyze the risk level and degree of safety hazards on roads, thus hindering effective prevention and targeted control measures. This increases the probability of traffic accidents and easily leads to them. Therefore, developing reasonable risk warning systems for roads has become one of the effective means to reduce traffic accidents.

[0003] Currently, the assessment methods for vehicles driving on slippery roads are mainly divided into:

[0004] (1) Static Infrastructure Assessment Method: The core principle of this method is to deploy fixed sensing devices at key road locations to directly or indirectly measure environmental and pavement physical parameters related to road hazards. It assumes that anomalies in these parameters (such as temperatures below freezing or the presence of a water film on the pavement) have a direct causal relationship with driving risks. Its advantages are that the data is direct, reliable, and continuous. However, its disadvantages are that it has a small coverage area, high cost, inflexibility, and cannot reflect vehicle behavior.

[0005] (2) Vehicle-based sensor-based assessment method: The core principle of this method is to use the vehicle's own perception system to monitor the vehicle's dynamic response and external environment, thereby identifying whether the vehicle is in or about to be in a dangerous state. It directly detects the "effect" of danger from the perspective of the "victim" (vehicle). This method can be implemented in real time, without requiring roadside modifications, and reflects the status of a single vehicle. However, this method has a limited perception range, relies on penetration rate, suffers from data silos, and has a delayed early warning.

[0006] (3) Road assessment method based on macroscopic statistical models: The core principle of this method is to use the statistical regularity of historical data, combined with real-time macroscopic weather forecasts, to predict which road sections have a high probability of risk in the future. It is a prediction method based on "historical repetition" and "physical law deduction". This method has the advantages of large-scale prediction and low cost, but it has poor accuracy, low real-time performance, and lacks microscopic insight.

[0007] In summary, existing slippery road risk assessment systems still have some shortcomings. First, current road risk assessment methods require the deployment of sensing devices on roads or vehicles to extract relevant environmental and vehicle data, which is relatively costly and highly susceptible to weather and natural environmental factors. Their predictive range is also relatively limited, exhibiting certain limitations (unable to address risk warnings for multiple hazardous road sections), making them unsuitable for widespread promotion and use. Second, existing risk assessment methods face difficulties in obtaining predictive parameters, require sophisticated sensing equipment, and are susceptible to poor sensitivity to external environmental factors, easily leading to false alarms. Summary of the Invention

[0008] In view of the shortcomings and deficiencies of existing technologies, this invention proposes a risk assessment method for slippery roads based on UAV vehicle trajectory recognition. This method employs a high-resolution UAV equipped with a gimbal to extract video data of vehicle trajectories on dangerous road sections, mapping these data to obtain vehicle ground trajectory sequences. Based on reference trajectory curves, it calculates parameter indicators for normal samples and slippery roads, compares the distribution differences between slippery roads and normal samples, and comprehensively considers indicators such as weighted overthreshold proportion, duration, and spatial consistency to obtain a risk score for vehicles driving on dangerous road sections. This approach significantly improves the accuracy, real-time performance, and interpretability of risk warnings, effectively reducing the false alarm rate caused by environmental misjudgments or equipment errors.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] A method for risk assessment of slippery roads based on vehicle trajectory recognition by unmanned aerial vehicles (UAVs) includes the following steps:

[0011] Step S1. Use the UAV to collect data and output the original video frames and flight control logs. Interpolate the meteorological data time series to the frame time to form a unified "frame-attitude-meteorological" time axis and generate one-time data collection.

[0012] Step S2. Perform camera calibration and ground mapping to construct at least one pixel-to-ground mapping model;

[0013] Step S3. Detect and track vehicles in the drone video in a time sequence, convert them from pixel coordinates to ground coordinates based on the pixel-to-ground mapping model, and output the vehicle ground trajectory sequence;

[0014] Step S4. Under normal weather conditions, calculate the parameter indicators of normal samples based on the reference trajectory curve, including lateral offset, heading error, angular velocity fluctuation, and velocity-direction coupling index. Calculate the trajectory oscillation frequency based on the acquired data, estimate the road surface adhesion coefficient and dynamic stability boundary, calculate the trajectory curvature anomaly index and yaw-sideslip coupling degree, and output all data after normalization to construct a normal weather distribution database.

[0015] Step S5. Based on the multi-vehicle trajectories and reference trajectory curves obtained in step S3, output standardized indicators that can be directly compared with the normal weather distribution database under slippery road conditions, and generate above-threshold indicators.

[0016] Step S6. Compare the standardized indicators output in step S5 with the normal weather distribution database in step S4 and output the distribution differences; adaptively adjust the above-threshold indicators under slippery roads based on meteorological coefficients, and calculate the weighted above-threshold ratio and the duration exceeding the 95th percentile of normal weather; evaluate the adjacency consistency and intra-segment spatial coherence, and finally form the segment-level risk score and risk level.

[0017] As a preferred embodiment of the present invention, after camera calibration in step S2, all aerial frames are first distorted and then ground mapping is performed. Based on the road conditions, a pixel-to-ground mapping model is established for flat road sections, and another pixel-to-ground mapping model is established for undulating / bridge / slope road sections.

[0018] As a preferred embodiment of the present invention, step S3 uses YOLOv8 to output vehicle candidate boxes and confidence scores for each frame, and then applies the DeepSORT multi-object tracking algorithm to achieve cross-frame vehicle ID matching, thereby determining the vehicle trajectory.

[0019] As a preferred embodiment of the present invention, in step S4, the target road segment is first divided into multiple segments along the main direction of the road. Based on the multi-vehicle trajectory obtained in step S3, a reference trajectory curve is fitted for each road segment on a normal weather sample. The fitting process converges in an iterative manner of "finding the nearest point, calculating the normal residual, updating parameters, and correcting the curve". After the fitting is completed, the curves of each segment are spliced ​​together to ensure that the segments are first-order continuous at the endpoints.

[0020] As a preferred embodiment of the present invention, in step S4, after fitting the reference trajectory curve for each segment, the scenario dimensions are further distinguished according to the curvature level and lane. A normal weather distribution database is constructed for each scenario dimension, and the slippery road sample data is compared with the normal weather distribution database according to the scenario dimension.

[0021] As a preferred embodiment of the present invention, in step S5, the trajectory oscillation frequency is calculated based on the lateral offset; the sideslip angle is calculated based on the speed, yaw rate, heading and trajectory direction to obtain the estimated road adhesion coefficient, and the adhesion estimation index and dynamic stability boundary are calculated; the trajectory curvature anomaly index is calculated based on the instantaneous curvature and expected curvature of the actual vehicle trajectory; and the yaw-sideslip coupling degree is calculated based on the yaw rate and the centroid sideslip angle.

[0022] As a preferred embodiment of the present invention, the risk score at time k in the m-th segment of step S6 The expression is:

[0023] ;

[0024] in, Specifically: lateral offset, heading error, angular velocity fluctuation, velocity-direction coupling, trajectory oscillation frequency, adhesion estimation index, trajectory curvature anomaly index, dynamic stability boundary, yaw-sideslip coupling degree; m: road segment; k: time window; ; Non-negative weights; Threshold shrinkage coefficient; Adjacency consistency; Spatial coherence within a segment; The expected or average duration for which a value continuously exceeds the 95th percentile of normal weather; The minimum cost required to transform an empirical distribution into a normal baseline distribution; : The proportion of the 95th and 99th percentiles; c is the bias term;

[0025] The risk level is then determined based on the risk score.

[0026] As a further preferred embodiment of the present invention, adjacency consistency ;

[0027] in, This is a preliminary score for the previous step or cycle. The weighted sum of the lateral offset and heading error, or the composite slip suspicion level. The mean value within the window, Represents segmentation, The total number of road segments. The threshold value is used.

[0028] As a further preferred embodiment of the present invention, the minimum cost required to convert the empirical distribution to the normal baseline distribution is... The expression is:

[0029] ;

[0030] in, The minimum cost required to transform the current window's empirical distribution into a normal baseline distribution. For the weighted experience distribution within the window, This represents the normal distribution of step S4.

[0031] The present invention also provides a computer-readable medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method for assessing the risk of slippery roads based on unmanned aerial vehicle (UAV) vehicle trajectory identification.

[0032] Advantages and beneficial effects of the present invention:

[0033] (1) The present invention adopts a high-resolution UAV equipped with a gimbal to extract video data of vehicle driving trajectory on dangerous road sections to obtain the risk score of vehicles driving on dangerous road sections, realizing non-contact, large-scale and dynamic road condition perception, and solving the problems of existing risk assessment methods that rely on fixed sensors, have small coverage, high deployment cost and poor flexibility.

[0034] (2) This invention uses the YOLOv8 depth detection model to realize vehicle recognition, applies the DeepSORT multi-target tracking algorithm to realize cross-frame vehicle ID matching, introduces ground calibration points and homography transformation, converts pixel trajectory into real-world coordinates, realizes high-precision and low-latency vehicle trajectory reconstruction, the whole process does not require environmental perception, and the parameters are easy to obtain, which solves the problems of existing risk assessment methods in data extraction, such as difficulty in obtaining parameters, high requirements for environmental perception, and susceptibility to weather and lighting.

[0035] (3) Due to the difference in road adhesion coefficient, the vehicle trajectory generated by the vehicle under normal weather and snowy weather is different for different dangerous road sections, provided that the speed limit is not exceeded. Based on this, the present invention constructs nine multi-dimensional trajectory difference indicators, including lateral offset, heading error, angular velocity fluctuation, velocity-direction coupling, trajectory oscillation frequency, adhesion estimation index, trajectory curvature anomaly index, dynamic stability boundary and yaw-sideslip coupling degree, to quantify the difference in trajectory behavior under snowy and normal weather, analyze its causes, and determine the risk level of road sections based on the difference analysis of multi-indicator fusion. This significantly improves the accuracy, real-time performance and interpretability of risk warning, and effectively reduces the false alarm rate caused by environmental misjudgment or equipment error.

[0036] (4) The present invention innovatively introduces the trajectory oscillation frequency (TOF) index, which can quantify the lateral vibration frequency of the vehicle on the low-adhesion road surface, capture the high-frequency instability behavior that traditional lateral offset index cannot reflect, and enhance the system’s sensitivity to dynamic slip state and early warning amount.

[0037] (5) By introducing the adhesion estimation index (AEI), this invention realizes the real-time adhesion coefficient inversion of the road surface based on vehicle trajectory dynamics, directly linking behavioral data with the physical state of the road surface, significantly improving the physical interpretability and engineering applicability of risk assessment, and providing a direct basis for maintenance decisions.

[0038] (6) The Trajectory Curvature Anomaly Index (TCAI) proposed in this invention can effectively identify the mismatch between the actual trajectory of the vehicle and the road geometry design caused by insufficient adhesion. It is particularly suitable for identifying slip risk in road sections with curvature changes such as curves and ramps, and makes up for the shortcomings of traditional indicators in geometric adaptability assessment.

[0039] (7) By constructing the Dynamic Stability Boundary (DSB) index, this invention realizes the quantitative assessment of instability proximity based on vehicle dynamics, which can provide early warning before the vehicle has obviously skidded, realizing the leap from "post-event judgment" to "pre-event warning", and providing key support for proactive traffic safety management.

[0040] (8) The introduced yaw-sideslip coupling index can characterize the coupling oscillation intensity of vehicle yaw and sideslip motion, and is particularly suitable for identifying the "fishtail" instability mode on icy and snowy roads, enhancing the system's ability to characterize complex instability mechanisms.

[0041] (9) This invention reduces reliance on external hardware facilities, requiring no modification to roads or vehicles. Risk assessment can be completed solely through drone aerial photography and video analysis, offering advantages such as rapid deployment, low cost, easy maintenance, and strong replicability. It is particularly suitable for road sections in mountainous areas, bridges and tunnels, and curved slopes where fixed monitoring is difficult to cover. This invention is not only applicable to icy and slippery road surfaces but can also be extended to other low-adhesion road surface scenarios such as water accumulation, oil stains, and gravel, demonstrating good scenario generalization ability and business extension potential, providing a unified technical framework for all-weather road safety monitoring. Attached Figure Description

[0042] Other objects and results of the invention will become more apparent and readily understood with reference to the following description taken in conjunction with the accompanying drawings. In the drawings:

[0043] Figure 1 The present invention provides a flowchart of a method for risk assessment of slippery roads based on vehicle trajectory identification by unmanned aerial vehicles. Detailed Implementation

[0044] To enable those skilled in the art to better understand the technical solutions and advantages of the present invention, the present application will be described in detail below with reference to the accompanying drawings, but this is not intended to limit the scope of protection of the present invention.

[0045] like Figure 1 As shown in the figure, this embodiment provides a method for risk assessment of slippery roads based on vehicle trajectory recognition by unmanned aerial vehicles. The method includes the following steps:

[0046] Step S1. Drone video preprocessing:

[0047] Before deployment, the scope of the aerial photography segment, road width, potential hazards (curves, bridges, long downhill slopes, shaded areas), and target resolution (GSD) were clearly defined. Based on the typical speed of vehicles on this road segment, the required frame rate and exposure limits were determined to ensure stable completion of the "detection-tracking-trajectory calculation" process.

[0048] Step S1.1. Preparation for UAV data collection:

[0049] The camera selected for this invention is: 4K@30–60fps, three-axis gimbal, global shutter; to reduce perspective distortion, a short-to-medium focal length (35–50mm equivalent) is used. Flight parameters are: altitude 60–120m, top-down or small-angle oblique view; shutter speed ≥1 / 1000s to suppress motion blur; in snowy weather, use low ISO and short exposure as much as possible, and add ND / polarizing filters to reduce flare if necessary. A top-down or small-angle oblique view position is used to ensure the field of view fully covers the road width and leaves 10–30% lateral redundancy. For long road sections, a single or serpentine flight path can be designed along the road direction; for key locations such as curves, slopes, and bridges, fixed-point observation (hovering) can be used to obtain continuous timing.

[0050] Step S1.2. Time and data synchronization:

[0051] Before takeoff, time reference unification is performed: the camera / flight controller is synchronized with UTC (Universal Time Coordinated) (hardware PPS (Pulse Per Second) or NTP (Network Time Protocol). During aerial photography, frame timestamps and flight control logs (attitude, altitude, track) are recorded, and prior information such as temperature, snowfall, and wind speed from the site or nearby weather stations is collected. If there are visual synchronization events (such as flashing lights or intersection signal switching), they are retained in the footage for subsequent correction of time offsets and drift.

[0052] Step S1.3. On-site sampling execution:

[0053] Select a weather window with wind speed and visibility that meets the SOP (Standard Operating Procedure) for flight. In stationary monitoring mode, keep the drone's ground speed as close to zero as possible to reduce background motion; in cruise mode, control the drone's speed to avoid excessively rapid image movement that could cause additional optical flow interference. Maintain the stability of the three-axis gimbal and minimize large changes in attitude. In nighttime or backlit conditions, prioritize adjusting the camera position and exposure strategy, and supplement with thermal infrared if necessary to improve detection stability.

[0054] Step S1.4. Data Processing and Output:

[0055] The original video and flight control logs are exported, and the timestamps of each frame are corrected based on the time synchronization results. Meteorological data time series are interpolated to the frame timestamps to form a unified "frame-attitude-meteorological" timeline, generating one-time data acquisition. Specifically, this includes: video sequences. With corrected frame time stamp Flight path and attitude (flight control log): Time-based meteorological sequence ; in, This represents the sequence of raw video frames acquired sequentially, where k is the frame index. This represents the image data of the k-th frame; Represents the Coordinated Universal Time (UTC) timestamp that corresponds precisely to each frame k; This represents the drone's flight altitude at frame k, usually referring to the vertical distance relative to the takeoff point or the ground, typically expressed in meters (m). The roll angle of the UAV at the k-th frame is the angle by which the fuselage rotates around its longitudinal axis, usually expressed in radians (rad) or degrees (°). The pitch angle of the UAV at the k-th frame is the angle of rotation of the fuselage around its horizontal axis, usually expressed in radians (rad) or degrees (°). The yaw angle of the UAV at the k-th frame is the angle (heading) of the fuselage rotating around its vertical axis, usually expressed in radians (rad) or degrees (°). For at frame time The estimated ambient temperature is usually expressed in degrees Celsius (°C). For at frame time The estimated intensity of precipitation / snowfall is usually expressed in millimeters per hour (mm / h). For at frame time The estimated wind speed is usually expressed in meters per second (m / s).

[0056] In this embodiment, the collected data is used as input for subsequent calibration and mapping (step S2) and vehicle detection and tracking (step S3).

[0057] Step S2. Camera calibration and ground mapping:

[0058] This step aims to reliably and consistently convert the pixel positions in the drone's imagery into road plane coordinates. Inputs include: video frames acquired in step S1, flight control attitude / altitude logs, ground control points (GCPs) or quantifiable road markings (such as zebra crossing spacing, lane width), and the road centerline established under normal weather conditions or the road reference trajectory interface to be learned later.

[0059] Step S2.1. Camera Calibration and Distortion Reduction:

[0060] Video captured by drones is affected by lens distortion and attitude changes, so camera geometric correction must be performed first. A pinhole imaging model is used to establish the projection relationship between pixel coordinates and real-space coordinates, and the camera's intrinsic and extrinsic parameters are obtained using a calibration board or ground control points.

[0061] In this embodiment, all aerial frames are first distorted and then ground mapping is performed.

[0062] Step S2.2. Ground mapping mode selection:

[0063] Automatic mode selection based on road section smoothness / superelevation / longitudinal slope and residual thresholds:

[0064] If the road segment is approximately planar and the reprojection error norm of the plane fitting is... The H-mode is used. If there are undulations / bridge deck camber / longitudinal slopes... If the planar residual exceeds the threshold, DSM orthorectification is used, i.e., DSM mode is used; if H mode Automatically switches to DSM mode; among which, Let be the reprojection error norm of the plane fitting; The residual threshold for the planar mode (H mode); This represents the elevation change rate (slope) of the road section. This is the slope threshold; This represents the average reprojection error; The average error threshold for H mode; This is due to scale bias; This is the scale deviation threshold. The unit is pixels, representing the smallest display unit in an image.

[0065] The data processing method for H mode (level road segment) is as follows:

[0066] Mark control points or line markers on the video frames to correspond one-to-one with the ground coordinate system, and register them to the real coordinate system. Select at least 4 ground control points (GCPs) or lane line corner points with known width / grid spacing to establish the correspondence between pixels and the ground. Use robust estimation (RANSAC, RANdom Sampling Consensus) to eliminate outliers and minimize the reprojection error to obtain the homography matrix.

[0067] The data processing method for DSM mode (undulating / bridge / slope road section) is: orthophoto projection / ray-surface intersection.

[0068] Select cross-view or multi-flight imagery for sparse / dense reconstruction to obtain camera pose and road segment DSM (regular grid or triangular network). Specifically, use multi-view frames or multiple camera positions to run SfM (Structure from Motion) to obtain the diluted point cloud outside the camera. Use MVS (Multi-View Stereo) to obtain the dense DSM (regular grid or triangular network TIN (Triangulated Irregular Network)). Then, for each pixel, find the intersection of the ray formed by the camera center and the pixel direction with the DSM to obtain the ground point corresponding to that pixel.

[0069] This embodiment can generate orthophotos in batches (unfolding the road surface onto a uniform horizontal plane), and subsequent detection / tracking can also be performed directly on the orthophoto sequence, reducing the calculation of intersection per pixel; the scale and ground resolution are checked with known ground scales (such as lane width and zebra crossing grid spacing) to form a consistency check.

[0070] Step S2.3. Preparation of the road reference system:

[0071] A road reference trajectory interface is prepared in the ground coordinate system: providing the ability to query the nearest point, tangent, and normal of any point to the reference trajectory. Downstream modules can then directly calculate key indicators such as "lateral offset (with sign)" and "reference heading," ensuring consistency across multiple videos and multiple user acquisitions.

[0072] In this embodiment, step 2 processes the input data to ultimately output distortion removal parameters and intrinsic parameters: Pixel-to-Ground Mapping Model: H (including) ) or DSM (grid / TIN); Pixel and ground API; H mode: DSM mode: Ray-to-surface intersection or orthophoto sequence; Road reference trajectory interface: Provides reference trajectory curves. Tangential / normal directions are used to support the calculation of the index in step S3; where: This is the camera intrinsic parameter matrix; The main radial distortion coefficient; For minor radial distortion coefficients; For higher-order radial distortion coefficients; Used to describe the tangential distortion in the x-direction; Used to describe the tangential distortion in the y-direction; These are the coordinates of the pixel on the ground. This represents the mapping relationship from ground coordinates backprojected to image pixel coordinates.

[0073] All of the above outputs are provided to subsequent vehicle detection and tracking, reference library (normal weather distribution library) construction, and difference analysis via standardized interfaces.

[0074] Step S3. Vehicle target recognition and temporal tracking stage:

[0075] This step converts the vehicles in the drone video into continuous, timestamped ground trajectories and attitude sequences, and outputs key behavioral indicators (velocity, heading, angular velocity, etc.) and confidence scores. Inputs include: the frame sequence processed in step S1 and a unified timescale, the pixel-to-ground mapping model (H or DSM) output in step S2, and the road reference trajectory interface.

[0076] Step S3.1. Robust preprocessing of the snow scene:

[0077] Each frame undergoes brightness and contrast stabilization to suppress overall graying under snowy conditions; isolated bright spots caused by snowfall are subject to temporal denoising or morphological cleanup; highlight suppression is enabled in strong reflective conditions, and polarizing filter shooting strategies are used when necessary to ensure clear vehicle edges and discernible textures.

[0078] A single-stage detection network (YOLOv8) is used to output vehicle candidate boxes and confidence scores for each frame, and overlapping boxes are removed through non-maximum suppression. Based on the scene and computing power, the detection threshold is set to a moderately conservative level, erring on the side of maximizing detection and minimizing misses. Optional appearance feature extraction (lightweight ReID head (Re-IDentification)) is enabled to provide feature support for subsequent re-association after occlusion. Due to the micro-motions and attitude adjustments of the drone, a set of image stabilization transformations is first obtained through inter-frame global motion estimation, transferring the detected vehicle center position to a "camera-stabilized coordinate system." The displacement after image stabilization better reflects the vehicle's true motion, which is beneficial for subsequent prediction and association.

[0079] Step S3.2. Pixel-to-ground coordinate mapping:

[0080] For each detected target, the midpoint of the bottom edge of the bounding box closest to the "grounding point" is preferentially selected as the representative position, and the mapping model provided in step S2 is used to convert it from pixel coordinates to ground coordinates. Homography mapping is used for smooth road sections, while DSM ray intersection or points are directly taken from the orthophoto to ensure consistent geometric scale for complex road sections such as undulating surfaces / bridge surfaces.

[0081] Step S3.3. Multi-target tracking (Kalman + Hungarian + ReID / IoU hybrid cost):

[0082] The standard process of "Kalman prediction + cost matrix matching (Hungarian algorithm) + trajectory management" is adopted:

[0083] For each on-orbit target, short-term motion prediction is performed to form the expected position and uncertainty of the next frame; a matching cost between detection and prediction is constructed, which integrates "motion consistency", "geometric overlap (IoU) (Intersection over Union)" and "appearance similarity", and completes one-to-one matching through the Hungarian algorithm; successfully matched trajectories are updated in state, unmatched trajectories are counted as missing, and deleted after consecutive missing trajectories exceed a threshold; newly generated targets must be continuously matched within several frames to be confirmed as valid trajectories, avoiding false trajectories caused by instantaneous false detection.

[0084] Ground speed / heading / angular velocity / acceleration calculation and robust handling:

[0085] Under a unified UTC timescale, velocity and heading are calculated based on the ground positions of adjacent frames, and derived quantities such as angular velocity and acceleration are obtained accordingly. To suppress jitter caused by noise and snow, a light smoothing (exponential smoothing or one-dimensional Kalman smoothing) is introduced, and the angle is unfolded / wound to avoid abrupt changes when crossing 0° / 360°.

[0086] Given trajectory ground sequence velocity vector ;in, Here are the ground coordinates (in meters) of the vehicle in the k-th frame. The time difference (in seconds) between the k-th frame and the (k-1)-th frame. Here is the velocity vector (in meters per second) for the k-th frame, and the heading angle for the k-th frame. ; Angle unfolding / winding treatment: ; This is the heading angle for the k-th frame after unfolding (to avoid angle jumps); it needs to be compared with the reference tangent. When comparing, wrap it back to [-π, π); angular velocity (yaw rate) of the k-th frame: Acceleration in the kth frame: .

[0087] Robust smoothing (Huber / Exponential smoothing): Or As the observations enter a one-dimensional Kalman smoothing, R / Q is adaptively estimated by EM. For smoothing coefficients, This is the smoothed value of the k-th frame. Angular velocity, For speed, It is acceleration.

[0088] Local parameters in the overhead road reference frame:

[0089] Based on the road reference trajectory interface provided in step S2, the system queries the nearest point, tangential direction, and normal direction of the vehicle relative to the reference trajectory at each time step, thereby obtaining core indicators such as signed lateral offset and heading error. Within the sliding time window, it also statistically analyzes angular velocity fluctuations, offset extremes, and durations, directly providing data for subsequent distribution difference analysis and risk scoring. Given the reference trajectory curve C(s) and tangential / normal direction in S2:

[0090] Closest point and tangent / normal: ;

[0091] Lateral offset (with symbols): ;

[0092] Heading error: ;

[0093] Window fluctuation (length W seconds): ;

[0094] in, Represents vehicle location; The reference trajectory curve (road centerline). for The first derivative (tangent vector); This is the arc length parameter corresponding to the nearest point of the vehicle's position on the reference trajectory; For reference trajectory in The unit tangent vector at that point; For reference trajectory in The unit normal vector at that location, unit tangent vector Components in the x and y directions; For reference trajectory in The tangential angle at that point (also known as the reference heading) ) () represents the angle wrapping function, which restricts the angle to the range of [-π, π]; W is the sliding time window (e.g., 5~10 seconds); The number of samples within the window; This represents the i-th angular velocity value within the window. The average angular velocity within the window, indicated by the superscript. Represents the transpose symbol.

[0095] The above data will be used for distribution comparison and scoring in steps S5–S6.

[0096] Furthermore, this embodiment outputs multi-dimensional confidence scores for each trajectory, including: average detection score, motion gating margin, effective length, and missing rate; it also generates an estimate of the ground coordinate covariance. Segments with low confidence scores or high uncertainty will be automatically downweighted or used only for statistical reference to avoid increasing false alarms.

[0097] Single-frame measurement confidence: detection score (representing the confidence score of the j-th detected target in the k-th frame) and Mahalanobis distance gating margin.

[0098] Trajectory-level confidence: ;

[0099] in, The average score of the detection boxes; This represents the average value of the Mahalanobis distance gating margin; The effective trajectory length; miss _rate represents the trajectory missing rate; These are the weighting coefficients; () is the Sigmoid function, which compresses the output to (0,1).

[0100] Error propagation to ground coordinates: by S2 With Kalman The superposition gives the total covariance The two are superimposed using a linear approximation of error propagation, and the total covariance expression is:

[0101]

[0102] in, Let Jacobian matrix be the ground mapping function in k frames. The position sub-block of the posterior covariance matrix of the vehicle position state estimation by the Kalman filter in step S3; This uncertainty stems from the pixel-to-ground coordinate transformation in step S2, and its magnitude depends on the camera calibration accuracy, the distribution of ground control points, and the fitting residuals of the homography / DSM model. If the value is too large, the trajectory segment will only be used for low-weight statistics.

[0103] In this embodiment, step 3 processes the input data to ultimately output a set of trajectories. Each trajectory contains All results are presented in a standardized data structure for use in steps S4 to S6, eliminating the need for repeated coordinate transformations.

[0104] Step S4. Construction of the reference trajectory and distribution library:

[0105] This step learns a reference trajectory and behavior distribution library for roads on normal weather samples, providing a stable geometric and statistical baseline for differential analysis under icy and snowy conditions. The inputs are: the multi-vehicle trajectories (ground coordinates, speed, heading, angular velocity, confidence level) output from step S3, the road spatial extent, and optional coarse centerline or road network information.

[0106] Under normal weather conditions, step S3 outputs the multi-vehicle trajectory: ;in, These are the ground coordinates, timestamp, heading angle, angular velocity, velocity, and confidence level of the i-th trajectory, respectively.

[0107] Road segment length It is recommended to travel 50-200m.

[0108] Step S4.1. Segmentation, cleaning, and weighting of normal trajectory samples:

[0109] Divide the target road into equal-length segments (50-200m recommended, 100m commonly used) along its main direction. Create a unique ID and spatial bounding box for each segment. All subsequent trajectory points are merged into the corresponding segment based on their coordinates, forming "segment-level sample buckets" for convenient parallel processing and difference statistics. Divide the target road segment along its main direction into... Each segment is approximately L in length; if a centerline is not yet available, the principal direction u of the entire segment can be determined using linear regression / principal component analysis (PCA), and then projected coordinates can be used. Perform initial segmentation. is the ground coordinate vector of the trajectory point.

[0110] For trajectory points within each segment, weights are applied according to the confidence levels generated in step S3; robust rules are used to eliminate or reduce the weights of outliers, isolated points, and location jumps; representative points are retained for high-frequency repetitions within the same vehicle and time window to reduce computational redundancy. This step ensures that the fitting is not dominated by random noise.

[0111] Specifically, confidence weighting: assigning a weight to each point. ,in, The final weight of the i-th trajectory point; The original confidence level of the i-th trajectory point; The confidence level amplification factor has a value of 5-10.

[0112] Robust denoising (MAD) is performed on each initial lateral outlier segment:

[0113]

[0114] Where MAD is the adjusted median absolute deviation; x is the sample value to be detected; The median value of the sample; This represents the absolute median deviation. This represents the absolute deviation of the original median. The threshold for outlier detection is usually set to [value missing]. .

[0115] Lane clustering: When the road has multiple lanes or obvious separation between up and down lanes, the point cloud is clustered within the segment, and a reference trajectory is independently built within each lane cluster. If only the "center reference line of the entire segment" is needed, clustering can be skipped, and the data can be fitted directly to all points. The purpose of clustering is to avoid centerline drift caused by the mixing of multiple lanes.

[0116] In each section Pair set For DBSCAN / Mean-Shift (Density-Based Spatial Clustering of Applications with Noise) clustering, the distance metric can be: ;

[0117] in, The ground positions of two points (e.g., (X,Y) coordinates, in meters). Represents the set of real numbers; The heading angle (orientation / direction of the vehicle, preferably in radians) and roll angle are for the corresponding positions. Let be the Euclidean distance between two points (in meters), where is its square (in meters). 2 ; The minimum directed angle between two angles, already "angle wrapped" and aligned to... ; These are the weighting coefficients for angle differences. This yields the lane clusters. If only the overall centerline is needed, this step can be skipped.

[0118] Step S4.2. Reference trajectory fitting (spline + robust constraints):

[0119] Within each segment (or each lane cluster), a smooth spline is fitted to the road center reference line, and normal residual minimization and robust loss are used to suppress the influence of outliers. The fitting process converges iteratively using a "nearest point—parameter update—curve correction" approach, ensuring geometric smoothness without excessively smoothing out the true curvature changes. After fitting, the segmented curves are stitched together to ensure first-order continuity between segments at their endpoints.

[0120] With section Interior point set Fitting reference trajectory curve (The arc length parameter s increases). Let S be the lateral and longitudinal coordinates of the trajectory at point s; minimize the normal residual using constrained splines, and initialize the correspondence (first nearest point projection):

[0121] For each point First find the nearest point on the reference trajectory curve (or thick center line). Calculate the normal residual by considering the tangent / normal direction.

[0122] Normal residual (lateral deviation of a point from the curve perpendicular to the direction of travel) and objective function:

[0123] Normal residual: ;in, These are the position coordinates of the curve at the parameter s. Represents the nearest point The direction of the law.

[0124] Minimize the weighted robust objective (Huber) and add smoothing regularization: ;

[0125] Huber robust loss function: ;

[0126] in, Let be the normal residual from the i-th trajectory point to the reference trajectory; The weight of the i-th trajectory point; Huber robust loss function; The Huber loss threshold; For smoothing regularization coefficients; The second derivative (curvature change) of the reference trajectory.

[0127] Iterative optimization (ICP (Iterative Closest Point) style): Alternately perform nearest point projection. Update and spline parameter update, until (Maximum residual: the residual with the largest absolute value) converges with the objective function. If a reference trajectory for the center of each lane is required... In the corresponding cluster The same process is executed independently on top.

[0128] Tangential, normal, and curvature are calculated to obtain the fitted reference trajectory, forming a road reference system interface: any ground point can query its nearest position on the reference trajectory, along with its corresponding tangential and normal directions. This interface will be used later for the standardized calculation of indicators such as lateral offset and heading error. Curvature: Segment-level curvature statistics: Or take the maximum curvature ; , They are respectively , The first derivative with respect to the arc length parameter s, , for , The second derivative with respect to the arc length parameter s, For the m-th segment of the reference trajectory The total arc length, i.e., the curve length of the reference trajectory segment from the starting point to the ending point. The calculated segment-level curvature statistics. and This will serve as a key road geometry feature, used to define the scenario dimension of 'curvature level'. In subsequent steps S5 and S6, the indicators of the ice and snow samples will be compared and scored with the normal weather distribution database of the corresponding level based on the curvature level of the road segment they are located in. This ensures that road segments with different geometric shapes, such as curves and straight sections, are assessed using differentiated risk assessment standards, avoiding misjudgments caused by geometric differences.

[0129] Step S4.3. Frenet Projection and Benchmark Indices:

[0130] The trajectory points under normal weather conditions are projected onto the reference trajectory, and basic behavioral quantities such as signed lateral offset, heading error, and angular velocity window fluctuation are uniformly calculated in the road reference system. Velocity-direction coupled indicators can be supplemented according to time windows. All indicators are bound to the index of "segment-time window-scenario" to ensure the location accuracy of subsequent comparisons.

[0131] For each normal sample point Calculate the nearest point of trajectory point pi on the reference curve C(s).

[0132] Lateral offset (with symbols): ;in, Is the curve in the parameter The location coordinates at that point Represents the nearest point The direction of the law.

[0133] Heading error: ; in, For reference trajectory in Tangential angle at the location;

[0134] Angular velocity / window fluctuation (given by S3) (Statistics within window W) ;

[0135] Velocity-direction coupling index (example) ;

[0136] in, Let velocity v and angular velocity be the angular velocity. In the time window Covariance within, For velocity v in the window within the standard deviation, angular velocity In the window within the standard deviation, For heading angle In the window within the standard deviation, For velocity v in the window The average value within, To prevent the division by zero, a small constant is usually taken as... .

[0137] Distributed library Construction: For each segment, further differentiate scenario dimensions as needed (e.g., lane, curvature level, speed level, road type). Construct a weighted empirical distribution for the above indicators, and save the median and key quantiles (P85, P90, P95, P99) and robust scaling statistics (mean / variance or median / MAD). When the sample size is insufficient, only the necessary quantiles are retained to reduce overfitting, for each segment... Establish empirical distributions for the following indicators respectively:

[0138] Weighted empirical distribution and quantiles: cumulative weights after sorting Total weight Weighted quantiles : Take the minimum (The index value of the j-th sample) makes Robust mean / variance: Alternatively, the median plus MAD can be used as a robust metric.

[0139] Percentage normalization and standardized output that can be directly called:

[0140] To facilitate subsequent scoring of differences, each indicator is normalized using the "median-high quantile" to generate a dimensionless z-shaped score. This normalization is calculated separately for each "segment / scenario" dimension to ensure comparability across different roads and geometric shapes. A quantile-normalized z-score is defined for each indicator x in the segment / scenario dimension: ;

[0141] in, For indicators to be normalized (such as lateral offset, angular velocity fluctuation, etc.); This represents the median of the indicator under normal weather conditions; This represents the 95th percentile of the indicator under normal weather conditions; It should be a very small constant to prevent division by zero. It should also prevent near-zero division (e.g., ...). ×Dimensional factor). For extremely steady-state road sections, it can be replaced with... Alternative .

[0142] Step S4.4. Self-check for consistency between reference trajectory and distribution:

[0143] After completing the fitting and distribution statistics, three self-checks are performed: Geometric consistency: the projection residuals are within a reasonable range and there are no abnormal peaks at the connection between segments; Scale consistency: the length of the reference trajectory matches the known scale of the ground; Statistical stability: the index distribution has no abnormal multi-peaks or extremely sparse samples.

[0144] Root Mean Square (RMS) of reprojection: Smoothing regularity: It should not increase abnormally; Scale deviation: checked against known lengths on the ground, error .

[0145] in, The weighted root mean square error of the normal reprojection; The weight of the i-th trajectory point; Total weight; This is for integrating the smooth regularized term.

[0146] If any one of the criteria is not met, prioritize refining the segmentation or enabling lane clustering; if still insufficient, revert to a coarser scenario level (e.g., retain only the segment-level distribution) or use migration from adjacent segments to supplement.

[0147] In this embodiment, step 4 processes the data to ultimately output a reference trajectory set: or ,as well as Segment-level statistics Distributed library: Stores data for each segment / scenario. Projection interface: Quality Report: , total sample weights W, multimodal indicators (e.g.) ) etc.; among which: : The reference trajectory curve of segment m, : The first segment of road m The reference trajectory curve of the lane; Reference heading angle; Curvature; Segment-level curvature statistics; Mean; :variance; :Percentile; : The coordinates of any ground point; Assume the arc length of the nearest point to point p on the reference trajectory; : The signed lateral offset of point p relative to the reference trajectory; Assuming point p has a heading angle, output the error between it and the reference heading.

[0148] Step S5. Calculation and standardization of wet and slippery road surface sample indicators

[0149] This step, under smooth road conditions, performs reference frame alignment, index calculation, window statistics, and standardization on the vehicle ground trajectories output by S3, producing dimensionless results that can be directly compared with the normal weather distribution database. Inputs include: a sequence of wet road surface sample trajectories (ground coordinates and unified time scale); a reference trajectory and interface (from S4): tangent / normal query; time window length W; and overlap step size. (Seconds, usually equal to the sampling interval), Normal Weather Distribution Database (from S4): Given for each segment / scenario Corresponding to Indicators such as SDI.

[0150] Step S5.1. Unifying the timeline and robust preprocessing

[0151] The trajectories of wet and slippery road surface samples (ice and snow samples) are resampled at a uniform time step to repair a small number of missing frames; weights are set for each time step based on the trajectory confidence given by S3; robust amplitude limiting or invalidation is performed on suspected jump points to avoid single-point anomalies affecting subsequent speed, heading and angular velocity calculations.

[0152] Resampling: Default Linear interpolation To equal-interval timestamps, amplitude limiting and skipping (based on MAD): for first-order difference . ,Exceed Jump point limiting or invalidation. Confidence level to weight: .

[0153] Step S5.2. Frenet Projection and Reference Frame

[0154] Closest point: ;

[0155] Signed lateral offset: ;

[0156] Reference heading: ;

[0157] Heading angle: ;

[0158] Heading error: ;

[0159] Angular velocity (yaw rate): ;

[0160] in, Let be the heading angle of the k-th frame. The heading angle is the heading angle of the (k-1)th frame.

[0161] Window statistics and volatility indicators: a sliding window of length W Calculation: Angular velocity fluctuation: Offset extremes and duration: Velocity-direction coupling: The above window statistics will be used for alignment and standardization with the distribution database of normal weather.

[0162] Smoothing and Robustnessing: To suppress jitter introduced by snowfall / occlusion, light smoothing is applied to key timing parameters: one-dimensional Kalman smoothing: As an observation, the process noise / measurement noise (Q,R) can be adaptively estimated using EM.

[0163] Alignment with normal distribution database (segment / scenario index): Find the road segment where each time k is located. With contextual dimensions (such as lane / curvature layer / speed layer), from (Indicates road segmentation) and contextual dimensions Read the quantiles of the corresponding indicators from the "Normal Weather Behavior Distribution Database" below:

[0164] .

[0165] Step S5.3. Quantile Normalization and Standardized Output

[0166] The various indicators obtained from the ice and snow samples in window statistics and fluctuation indicators are normalized according to the median and high quantile of their respective "segment / scenario," outputting dimensionless standardized scores to ensure comparability between different roads and different geometric characteristics. Simultaneously, "exceeding" [the specified threshold] is generated. or more The Boolean instruction for S6 statistical above-threshold proportions and persistence. For any indicator (For example The dimensionless z-score is obtained by normalizing using median-high quantiles; simultaneously, the suprathreshold index is output. Used for subsequent statistical analysis of "upper threshold proportion".

[0167] Composite slippage suspicion: The standardized sub-indices are combined linearly / robustly to obtain the time-level slippage suspicion. It can be used as an auxiliary indicator for real-time monitoring, or for calculating adjacency consistency in subsequent steps. Preliminary score:

[0168]

[0169] in: : Standardized value of lateral offset; Standardized value of heading error; Standardized values ​​of angular velocity fluctuations; Standardized value of the velocity-direction coupling index; The weights of each indicator satisfy the following: It can be determined by feature importance or validation set fitting.

[0170] Confidence and validity control: Sample confidence: ,in: For time window The percentage of missing items within; The basic weight of a single point at time k; Time window Effective sample rate within; : represents the weighting coefficients of the linear combination; : The combined confidence level of time k and index x. Invalid handling: When Or the number of valid samples in the window is insufficient (e.g., <70%). The confidence threshold is the value at that moment. Marked as NaN, only used for low-weighted statistics or direct exclusion. Uncertainty estimation: If necessary, estimate the ground coordinate covariance of S2. With S3 Kalman covariance P propagation to Above (linear approximation) ), used for subsequent robust aggregation in S6.

[0171] Step S5.4. Introduction and Calculation of Innovation Indicators

[0172] To further improve the accuracy and physical interpretability of road risk warnings under icy and snowy weather, this step introduces the following five innovative evaluation indicators based on existing standardized indicators (lateral offset, heading error, angular velocity fluctuation, and velocity-direction coupling index):

[0173] Trajectory Oscillation Frequency (TOF): The frequency at which a vehicle oscillates laterally per unit time, used to quantify the dynamic stability of a vehicle on low-adhesion icy or snowy surfaces.

[0174] The signed lateral offset sequence of the vehicle within the time window W Detrending processing is performed to obtain an oscillating signal with zero mean. .right Perform a Fast Fourier Transform (FFT) to extract the dominant frequency component:

[0175]

[0176] in, The amplitude of the spectrum; For frequency, The dominant frequency component reflects the frequency of the vehicle's lateral sway.

[0177] To suppress noise interference, only frequency components with amplitudes exceeding 50% of the maximum amplitude are retained, and the highest frequency among them is selected as TOF. TOF can capture the high-frequency vibration behavior unique to vehicles on icy and snowy roads, which is a dynamic instability characteristic that traditional lateral drift indicators cannot reflect.

[0178] Adhesion Estimation Index (AEI): A real-time estimate of the road surface adhesion coefficient obtained by back-calculation based on vehicle trajectory dynamics, used to directly characterize the risk of road slippage.

[0179] Adhesion coefficient inversion method: Based on the above monorail dynamics model, the tire lateral stiffness and road adhesion coefficient are inverted. Association, that is, let Establish the transverse dynamic equation:

[0180]

[0181] in, For lateral acceleration, For the lateral stiffness of the front and rear wheels, These are the front and rear wheel slip angles, where m represents the vehicle's mass. These are the nominal lateral stiffness of the front wheel and the nominal lateral stiffness of the rear wheel, respectively: representing the magnitude of the lateral force that the front (rear) wheel can provide for each 1 radian of lateral slip angle on a high-adhesion road surface (such as dry asphalt).

[0182] Estimate the sideslip angle using trajectory data:

[0183] in: The front wheel angle is calculated from the heading and trajectory direction. For longitudinal and lateral velocities, The yaw rate is angular velocity. The distance from the vehicle's center of gravity to the rear axle. : The distance from the vehicle's center of gravity to the front axle.

[0184] To further enhance the physical interpretability of risk warnings for icy and snowy roads, this method introduces an adhesion estimation index (AEI) based on vehicle dynamics, in addition to traditional trajectory indices, to directly characterize the real-time adhesion capability of the road surface.

[0185] The equivalent adhesion coefficient is back-calculated in real time using weighted least squares or Kalman filtering. And normalized to AEI:

[0186] ,

[0187] in, : The real-time adhesion coefficient obtained by reverse calculation; : is the typical adhesion coefficient of dry asphalt pavement (taken as 0.8~1.0).

[0188] AEI directly links vehicle trajectory behavior with road surface physical conditions, possessing clear physical meaning, which can significantly improve the interpretability and reliability of risk warnings.

[0189] Trajectory Curvature Anomaly Index (TCAI): The relative deviation between the instantaneous curvature of the actual vehicle trajectory and the expected curvature of the road geometry design. It is used to quantify trajectory geometry mismatch caused by insufficient adhesion.

[0190] Based on the road reference trajectory C(s) provided by S2, the desired curvature is calculated:

[0191]

[0192] Where s is the arc length parameter of the vehicle projected onto the reference trajectory at the current moment.

[0193] Vehicle ground trajectory sequence based on S3 output Calculate the instantaneous curvature of the actual trajectory :

[0194]

[0195] in, For heading angle, This represents the vehicle's ground coordinates in the k-th frame.

[0196] Calculate the trajectory curvature anomaly index :

[0197]

[0198] in: The instantaneous curvature of the vehicle's actual trajectory; : Desired curvature for road design; : The projected arc length of the vehicle on the reference trajectory; It is a small constant to prevent division by zero.

[0199] Dynamic Stability Boundary (DSB): A stability criterion based on vehicle dynamics theory, which quantitatively characterizes the relative distance between the vehicle's current state and the sideslip or fishtailing instability boundary.

[0200] Obtain real-time vehicle status data:

[0201] Longitudinal and lateral velocities (obtained by S3 trajectory sequence difference and coordinate transformation).

[0202] : Yaw angular velocity (S3 has been calculated).

[0203] Obtain the estimated road surface adhesion coefficient Calculate DSB:

[0204]

[0205] Where L is the vehicle wheelbase (which can be set to a typical value or a lightweight estimation algorithm); g is the acceleration due to gravity. Longitudinal and lateral velocities; : Yaw rate; : Estimate the adhesion coefficient.

[0206] Early warning mechanism: When DSB < 0.3, the vehicle is considered to be approaching the dynamic instability boundary, triggering a high-risk warning.

[0207] Yaw-Sideslip Coupling (YSC): The product of the temporal correlation coefficient between the yaw rate and the sideslip angle of the center of mass and their respective oscillation intensities, used to characterize the coupled oscillation intensity of the vehicle's yaw and sideslip motion.

[0208] Estimation of the center-of-gravity sideslip angle: The center-of-gravity sideslip angle β is the angle between the vehicle's velocity direction and its heading. It can be estimated using the following formula: ;in, The vehicle heading angle (which can be smoothly estimated from the trajectory direction or estimated together as a state variable). These are the vehicle's longitudinal velocity (along the vehicle's orientation) and lateral velocity (perpendicular to the vehicle's orientation) at time k.

[0209] Within a sliding time window W, collect the yaw rate sequence { } and centroid side slip angle sequence { }

[0210] Calculate YSC:

[0211]

[0212] in: The correlation coefficient between yaw rate and sideslip angle of the center of mass, with a range of values ​​of [value missing]. ; Standard deviation of yaw rate; Standard deviation of the centroid sideslip angle.

[0213] Regarding the aforementioned innovation indicators, If the same indicators are used, perform quantile normalization and standardized output.

[0214] This step outputs a standardized data layer for S6, including: original / smoothed data. Window statistics: Standardization results: Supra-threshold indicators: Composite fraction: Trajectory oscillation frequency: TOF; Adhesion estimation index: AEI; Trajectory curvature anomaly index: TCAI; Dynamic stability boundary: DSB; Yaw-sideslip coupling: YSC; Confidence level All results are stored according to scenario indexes for direct aggregation and difference statistics in S6.

[0215] in: : Lateral offset, heading error, angular velocity, instantaneous velocity, and instantaneous acceleration at time k; In the window Angular velocity window fluctuations, extreme values ​​of lateral offset, duration of large offset, and velocity-direction coupling index; Lateral offset exceeding P95 indication and heading error exceeding P99 indication.

[0216] Step S6. Differential Distribution Analysis and Risk Warning

[0217] This step aggregates the standardized indicators and over-threshold indicators produced by S5 at the granularity of "segmentation × time window", compares them with the normal weather distribution database of S4 at the distribution level, and incorporates meteorological priors (temperature, snowfall / precipitation, wind speed, etc.) into the adaptive adjustment of thresholds or scores to form segment-level risk scores and risk levels.

[0218] Inputs include: standardized scores from the S5 output (lateral offset, heading difference, angular velocity fluctuation, velocity-direction coupling, trajectory oscillation frequency, adhesion estimation index, trajectory curvature anomaly index, etc.). and suprathreshold indicators ; quantile information of the segmented / scenario "normal weather" distribution database established by S4; meteorological elements (temperature, snowfall / precipitation intensity, wind speed, etc.). Segment-window data: in the segmented With time window The standardized indicator sequence within; and time weight Normal weather distribution database: [Data on] the same "scenario" Meteorological priors: temperature T, snowfall / precipitation intensity P, wind speed W.

[0219] Step S6.1. In-window cleaning and polymerization

[0220] For each "segment × time window", samples with confidence levels below the threshold are first removed. The remaining samples are then weighted according to their confidence levels to form the effective sample set for that window. If the number of effective samples is insufficient, the time window is marked as "low confidence, observation only" to avoid misjudgments caused by sparse data.

[0221] Only confidence level is retained The sample, For the duration of the large offset, let the weights... Effective weights sum: ,like (Default 200), this window is for reference only and will not trigger a warning.

[0222] Step S6.2. Measurement of Distribution Difference

[0223] For each behavioral indicator, an empirical distribution is established within the current window and compared with the baseline distribution of the same "segment / scenario" under normal weather conditions, outputting evidence of distribution differences (e.g., shape differences, location shifts, etc.). For each indicator... Collect samples within the window .

[0224] Earth Mover's Distance (EMD, 1-Wasserstein) method for measuring difference:

[0225]

[0226] in, The minimum cost required to transform the current window's empirical distribution into a normal baseline distribution. For the weighted experience distribution within the window, This represents the normal distribution of S4 (same segment / same scenario).

[0227] Define meteorological coefficient (The colder, more humid, and windier the environment, the lower the threshold.)

[0228]

[0229] Before scoring, use Adjusting the quantile threshold: ;in: , The adjusted value is the threshold.

[0230] Weighted above-threshold proportion (primary evidence): ;

[0231] Duration (the longest consecutive time exceeding P95 within the window)

[0232]

[0233] in: For an index x, the expected or average value is defined as the duration during which the value of x continuously exceeds the 95th percentile of normal weather within a segment m and a time window k. : based on meteorological coefficient Adjusted 95th percentile; : The index at time t.

[0234] Pre-scoring calculation: To ensure consistency across the computational space, a simplified pre-scoring is first calculated based on the core evidence of the current window. For example, it can be defined as the weighted upper-threshold sum of the lateral offset and the heading error:

[0235]

[0236] in: The pre-scoring is based on a preset weight (e.g., all values ​​are 0.5). This pre-scoring is only used for adjacency consistency assessment and does not participate in the final risk score. The calculation.

[0237] Adjacency consistency:

[0238] in, Pre-scoring for the previous step or cycle (if none, use instead) (Threshold judgment), can be used Represents segmentation, The total number of road segments. The threshold value is used.

[0239] Step S6.3. Stage Scoring and Risk Assessment

[0240] Construct segment-level feature vectors:

[0241]

[0242] Linear normalization + Sigmoid function:

[0243] in, Specifically, there are nine indicators: lateral drift, heading error, angular velocity fluctuation, velocity-direction coupling, trajectory oscillation frequency, adhesion estimation index, trajectory curvature anomaly index, dynamic stability boundary, and yaw-sideslip coupling degree. Segment-level risk score, range m: road segment; k: time window; Compress the linear combination to ; Non-negative weights; Threshold shrinkage coefficient; Adjacency consistency; Spatial coherence within a segment is approximated by the spatial autocorrelation index of the trajectory point risk value and the inverse of the risk variance within the sliding window. : Persistent score; Differences in Earth Mover's Distance distribution; : The proportion exceeding the 95th / 99th percentile; c is the bias term (constant term) used to adjust the baseline level of the risk score, which can be calibrated using historical data or set as an empirical value.

[0244] Risk classification:

[0245]

[0246] In this embodiment, different control measures are adopted according to different risk levels, as follows:

[0247] Level: Standard speed limit; keep all lanes open, do not apply pesticides; routine patrols of key points (bridge surface, ramps, shady slopes) every 1-2 hours.

[0248] Level: Speed ​​limit reduced by 20%–30% (e.g., from 100 km / h to 70–80 km / h); if necessary, close the outermost lane or one lane at the bend of a dangerous section, keeping the main lane open. Apply snow-melting / anti-skid agents (on bridge surfaces, ramps, curves, slopes, and shaded areas); broadcast warnings via VMS / broadcast: "Ice and snow risk section, reduce speed"; provide advance warning 2-3 times before the dangerous section (1500m / 800m / 300m).

[0249] Level: Speed ​​limit reduced by 40%–50% (e.g., 100 → 50-60 km / h); temporary single-lane traffic or alternating traffic flow (depending on road width and queue length); diversion / ramp closure if necessary, detour to cleared or treated sections. Increase the frequency of de-icing agent application (double coat in both directions, focusing on bends and bridge joints); combine mechanical de-icing / snow removal with sand application to increase friction; VMS / broadcast / navigation push notifications "High-risk section, slow down / may be closed"; establish fixed monitoring positions at key locations such as bends, slopes, and bridges for continuous assessment and real-time warnings.

[0250] The present invention also provides an electronic device, comprising: one or more processors and a memory; wherein the memory is used to store one or more programs, and when the one or more programs are executed by the one or more processors, the one or more processors implement the wet and slippery road risk assessment method described above.

[0251] The present invention also provides a computer-readable medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned method for assessing the risk of slippery roads.

[0252] Those skilled in the art will understand that all or part of the functions of the various methods / modules in the above embodiments can be implemented by hardware or by computer programs. When all or part of the functions in the above embodiments are implemented by computer programs, the program can be stored in a computer-readable storage medium, which may include: read-only memory, random access memory, disk, optical disk, hard disk, etc., and the above functions can be implemented by executing the program with a computer. For example, the program can be stored in the memory of a device, and when the program in the memory is executed by the processor, all or part of the above functions can be implemented.

[0253] In addition, when all or part of the functions in the above embodiments are implemented by computer programs, the programs can also be stored in storage media such as servers, other computers, disks, optical discs, flash drives, or portable hard drives. They can be downloaded or copied to the memory of the local device, or the system of the local device can be updated. When the program in the memory is executed by the processor, all or part of the functions in the above embodiments can be implemented.

[0254] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention 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 the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of protection described in the claims.

Claims

1. A method for assessing the risk of slippery roads based on unmanned aerial vehicle (UAV) vehicle trajectory recognition, characterized in that, The method includes the following steps: Step S1. Use the UAV to collect data and output the original video frames and flight control logs. Interpolate the meteorological data time series to the frame time to form a unified "frame-attitude-meteorological" time axis and generate one-time data collection. Step S2. Perform camera calibration and ground mapping to construct at least one pixel-to-ground mapping model; Step S3. Detect and track vehicles in the drone video in a time sequence, convert them from pixel coordinates to ground coordinates based on the pixel-to-ground mapping model, and output the vehicle ground trajectory sequence; Step S4. Under normal weather conditions, calculate the parameter indicators of normal samples based on the reference trajectory curve, including lateral offset, heading error, angular velocity fluctuation, and velocity-direction coupling index. Calculate the trajectory oscillation frequency based on the acquired data, estimate the road surface adhesion coefficient and dynamic stability boundary, calculate the trajectory curvature anomaly index and yaw-sideslip coupling degree, and output all data after normalization to construct a normal weather distribution database. Step S5. Based on the multi-vehicle trajectories and reference trajectory curves obtained in step S3, output standardized indicators that can be directly compared with the normal weather distribution database under slippery road conditions, and generate above-threshold indicators. Step S6. Compare the standardized indicators output in step S5 with the normal weather distribution database in step S4 and output the distribution differences; adaptively adjust the above-threshold indicators under slippery roads based on meteorological coefficients, and calculate the weighted above-threshold ratio and the duration exceeding the 95th percentile of normal weather; evaluate the adjacency consistency and intra-segment spatial coherence, and finally form the segment-level risk score and risk level.

2. The method for assessing the risk of slippery roads based on UAV vehicle trajectory recognition according to claim 1, characterized in that, In step S2, after the camera is calibrated, all aerial frames are first distorted and then ground mapping is performed. Based on the road conditions, a pixel-to-ground mapping model is established for flat road sections, and another pixel-to-ground mapping model is established for undulating / bridge / slope road sections.

3. The method for assessing the risk of slippery roads based on UAV vehicle trajectory recognition according to claim 1, characterized in that, Step S3 uses YOLOv8 to output vehicle candidate boxes and confidence scores for each frame, and then applies the DeepSORT multi-object tracking algorithm to achieve cross-frame vehicle ID matching, thereby determining the vehicle trajectory.

4. The method for assessing the risk of slippery roads based on UAV vehicle trajectory recognition according to claim 1, characterized in that, In step S4, the target road segment is first divided into multiple segments along the main road direction. Based on the multi-vehicle trajectory obtained in step S3, a reference trajectory curve is fitted for each road segment on a normal weather sample. The fitting process converges in an iterative manner of "finding the nearest point, calculating the normal residual, updating parameters, and correcting the curve". After the fitting is completed, the curves of each segment are spliced ​​together to ensure that the segments are first-order continuous at the endpoints.

5. The method for assessing the risk of slippery roads based on UAV vehicle trajectory recognition according to claim 4, characterized in that, In step S4, after fitting the reference trajectory curve for each segment, the scenario dimensions are further distinguished according to curvature level and lane. A normal weather distribution database is constructed for each scenario dimension, and the wet and slippery road sample data is compared with the normal weather distribution database according to the scenario dimension.

6. The method for assessing the risk of slippery roads based on vehicle trajectory recognition by unmanned aerial vehicles according to claim 5, characterized in that, In step S5, the trajectory oscillation frequency is calculated based on the lateral offset; the sideslip angle is calculated based on the speed, yaw rate, heading and trajectory direction to obtain the estimated road adhesion coefficient, and the adhesion estimation index and dynamic stability boundary are calculated; the trajectory curvature anomaly index is calculated based on the instantaneous curvature and expected curvature of the actual vehicle trajectory; and the yaw-sideslip coupling degree is calculated based on the yaw rate and the centroid sideslip angle.

7. The method for assessing the risk of slippery roads based on UAV vehicle trajectory recognition according to claim 6, characterized in that, Risk score at time k in segment m in step S6 The expression is: ； in, Specifically: lateral offset, heading error, angular velocity fluctuation, velocity-direction coupling, trajectory oscillation frequency, adhesion estimation index, trajectory curvature anomaly index, dynamic stability boundary, yaw-sideslip coupling degree; m: road segment; k: time window; ; Non-negative weights; Threshold shrinkage coefficient; Adjacency consistency; Spatial coherence within a segment; The expected or average duration for which a value continuously exceeds the 95th percentile of normal weather; The minimum cost required to transform an empirical distribution into a normal baseline distribution; : The proportion of the 95th and 99th percentiles; c is the bias term; The risk level is then determined based on the risk score.

8. The method for assessing the risk of slippery roads based on vehicle trajectory recognition by unmanned aerial vehicles according to claim 7, characterized in that, Adjacency consistency ; in, This is a preliminary score for the previous step or cycle. The weighted sum of the lateral offset and heading error, or the composite slip suspicion level. The mean value within the window, Represents segmentation, The total number of road segments. The threshold value is used.

9. The method for assessing the risk of slippery roads based on vehicle trajectory recognition by unmanned aerial vehicles according to claim 7, characterized in that, Minimum cost required to transform an empirical distribution into a normal baseline distribution The expression is: ； in, The minimum cost required to transform the current window's empirical distribution into a normal baseline distribution. For the weighted experience distribution within the window, This represents the normal distribution of step S4.

10. A computer-readable medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the slippery road risk assessment method based on UAV vehicle trajectory identification as described in any one of claims 1 to 9.

Images