Driving behavior monitoring methods, devices, and storage media based on dashcams

Abstract

This invention relates to the field of driving behavior monitoring technology, specifically to a driving behavior monitoring method, device, and storage medium based on a dashcam. It includes the following steps: acquiring video data from the dashcam and extracting image frames; performing multi-target detection on the image frames, identifying the position information and category labels of road participants in the image frames, and constructing the motion trajectory of road participants in the image sequence using a target tracking algorithm; simultaneously, using a curve perception trajectory correction method based on steering wheel angle to perform rotation compensation and correction on the motion trajectory; analyzing the spatial relative relationships and speed change trends between road participants based on the motion trajectory, and using an LSTM model to identify traffic movement events. This invention achieves high-precision tracking and correction of the motion trajectory of road participants by combining dashcam video data with the vehicle's own state information.

Description

Technical Field

[0001] This invention relates to the field of driving behavior monitoring technology, and more specifically, to a driving behavior monitoring method, device, and storage medium based on a driving recorder. Background Technology

[0002] In the current traffic environment, traditional dashcams are mainly used for accident liability determination, lacking the ability to monitor and analyze driving behavior in real time, making it difficult to effectively prevent traffic accidents. Especially when driving on curves, the dashcam's perspective changes due to vehicle turning, distorting the movement trajectories of road users in the image coordinate system. This leads to inaccurate analysis of spatial relationships (such as trends in distance changes) and speed changes, easily causing misjudgments or omissions in traffic incident assessments. Furthermore, existing technologies are insufficient in assessing drivers' behavioral responses to sudden traffic events, failing to comprehensively and scientifically evaluate whether drivers have made timely and appropriate reactions. These problems collectively result in inaccurate risk assessment of driving behavior, affecting driving safety and the effectiveness of accident prevention. Therefore, this paper proposes a driving behavior monitoring method, device, and storage medium based on dashcams. Summary of the Invention

[0003] The purpose of this invention is to provide a driving behavior monitoring method, device, and storage medium based on a dashcam, in order to solve the problems mentioned in the background art, such as the inability of existing dashcams to monitor driving behavior in real time, trajectory distortion during curve driving leading to misjudgment or missed judgment of traffic incidents, and inaccurate driver response assessment.

[0004] To achieve the above objectives, a driving behavior monitoring method based on a dashcam is provided, comprising the following steps:

[0005] S1. Collect video data from the dashcam and extract image frames;

[0006] S2. Perform multi-target detection on the image frame, identify the position information and category label of road participants in the image frame, and construct the motion trajectory of road participants in the image sequence by combining the target tracking algorithm. At the same time, use the curve perception trajectory correction method based on steering wheel angle to perform rotation compensation and correction on the motion trajectory.

[0007] S3. Analyze the spatial relative relationships and speed change trends among road participants based on motion trajectories, and use the LSTM model to identify traffic motion events;

[0008] S4. Acquire and synchronize the vehicle's own status data, and combine it with the identified traffic motion events to determine whether the driver has generated corresponding driving behavior;

[0009] S5. Establish a driving risk index to assess the risk of driving behavior, and make punitive adjustments to the driving risk index based on the driver's response delay time.

[0010] As a further improvement to this technical solution, in step S2, multi-target detection is performed on the image frame to identify the location information and category labels of road participants in the image frame, including the following steps:

[0011] S2.1 Perform preprocessing operations on the image frames;

[0012] S2.2 Analyze the preprocessed image frames based on the multi-target detection algorithm to identify road participants in the image;

[0013] S2.3 Output the bounding box coordinates, class probability distribution, and confidence score of each detected target road participant from the multi-object detection algorithm.

[0014] As a further improvement to this technical solution, the step of constructing the motion trajectory of road participants in the image sequence by combining the target tracking algorithm includes the following steps:

[0015] S2.4 In each frame of the image, identity management is performed on each detected target output by the multi-target detection model. If the target appears for the first time, a unique tracking identifier is assigned to it, and the corresponding tracker instance is initialized to record the target's state information and historical trajectory.

[0016] S2.5 For each detected target, crop an image patch within its bounding box region and input it into a deep neural network to extract appearance feature vectors;

[0017] S2.6. Use the target tracking algorithm to match the target detected in the current frame with the existing tracking trajectory in the previous frame, and update the state of each tracked target by combining the target position predicted by the Kalman filter.

[0018] S2.7 Record the position information of each target with a tracking identifier in each frame of the image in chronological order, and construct a temporal trajectory data structure;

[0019] S2.8. Using a curve perception trajectory correction method based on steering wheel angle, the time-series trajectory data structure is rotated and corrected.

[0020] As a further improvement to this technical solution, in step S2.8, a curve perception trajectory correction method based on steering wheel angle is used to perform rotation compensation and correction on the time-series trajectory data structure, including the following steps:

[0021] S2.81. Real-time synchronous acquisition of the vehicle's steering wheel angle signal via the vehicle's OBD-II interface;

[0022] S2.82. Based on the steering wheel angle and vehicle speed, calculate the vehicle's yaw rate to reflect the vehicle's current turning degree and rotation trend;

[0023] S2.83. Extract the image coordinate information of each tracked target from the current image frame and its historical image frames;

[0024] S2.84. For each historical trajectory point, calculate the time difference between it and the current image frame, and combine it with the yaw rate at the corresponding moment to calculate the cumulative change in the vehicle's rotation angle within that time difference.

[0025] S2.85. Rotate the coordinates of the target point in the image in the opposite direction by an angle to obtain the corrected trajectory coordinates. Based on the camera calibration parameters of the dashcam, map the corrected trajectory coordinates to the physical coordinates in the vehicle coordinate system. Replace the historical trajectory coordinates of each target with the mapped trajectory coordinates to construct the final time-series trajectory data structure.

[0026] As a further improvement to this technical solution, in step S3, the spatial relative relationships and speed change trends among road participants are analyzed based on the motion trajectory, and traffic motion events are identified using an LSTM model, including the following steps:

[0027] S3.1 For each pair of related targets, calculate the Euclidean distance and relative angle frame by frame based on their trajectory points;

[0028] S3.2 For each target, calculate the velocity vector at consecutive time points, calculate the relative velocity vector, and determine the direction and magnitude of the relative velocity;

[0029] S3.3 For each target, construct a trajectory-velocity-relative relationship sequence within a time window TN;

[0030] S3.4 Input the trajectory-speed-relative relationship sequence into the pre-trained LSTM model and output the corresponding traffic event category prediction.

[0031] As a further improvement to this technical solution, in step S4, the vehicle's own state data is acquired and synchronized, and combined with the identified traffic movement events, it is determined whether the driver has engaged in corresponding driving behavior, including the following steps:

[0032] S4.1 Acquire and synchronize the vehicle's own control status data, align it with the identified traffic motion events through a unified timestamp, and define the driver's response behavior for each type of traffic event;

[0033] S4.2 Setting the response judgment time window The system analyzes changes in vehicle control signals within a given time interval, using the central time point of traffic incident identification as a benchmark.

[0034] S4.3 Output the judgment result of the driver's response behavior to traffic movement events.

[0035] As a further improvement to this technical solution, in step S5, a driving risk index is established to assess the risk of driving behavior, including the following steps:

[0036] S5.1 Extract risk factors from identified traffic incidents and driver response behaviors;

[0037] S5.2 Normalize each risk factor to a uniform numerical range;

[0038] S5.3 By constructing a physical model of risk field strength, the risk field strength is calculated based on collision time, dynamic equivalent safety distance and environmental and regional correction factors, and a punitive correction is made in combination with driver response delay to dynamically calculate the driving risk index.

[0039] S5.4. Based on the driving risk index, driving behavior is divided into multiple risk levels;

[0040] S5.5 Output the risk assessment results.

[0041] As a further improvement to this technical solution, in S5.3, the driving risk index is calculated using a risk field strength physical model, involving the following specific steps:

[0042] S5.31. Based on the vehicle dynamics model, combined with the current speed of the vehicle, the relative speed between the vehicle and the target, the environmental friction coefficient and the driver's reaction time, calculate the minimum safe braking distance under the current environment.

[0043] S5.32. Predict the time required for a collision to occur while the vehicle is maintaining its current operating state by using the spatial distance and relative speed between the target and the vehicle.

[0044] S5.33. Based on the spatial field modeling method, and by introducing the regional risk multiplier and collision time decay factor, calculate the risk field intensity at the current moment;

[0045] S5.34. Adjust the penalty based on the driver's response behavior and output the final driving risk index.

[0046] On the other hand, the present invention provides a driving behavior monitoring device based on a driving recorder, comprising:

[0047] The acquisition module is used to acquire video data from the dashcam in front of the vehicle and extract image frames;

[0048] The analysis module is used to perform multi-target detection on image frames, identify the location information and category labels of road participants in the image frames, and construct the motion trajectory of road participants in the image sequence by combining the target tracking algorithm. At the same time, the motion trajectory is rotated and corrected by using the curve perception trajectory correction method based on steering wheel angle.

[0049] The traffic incident detection module is used to analyze the spatial relative relationships and speed change trends among road participants, and to identify traffic motion events using a traffic incident detection model.

[0050] The driving behavior recognition module is used to acquire and synchronize the vehicle's own status data, and combine it with the identified traffic movement events to determine whether the driver has generated the corresponding driving behavior.

[0051] The risk assessment module is used to establish a driving risk index to assess the risk of driving behavior, and to make punitive adjustments to the driving risk index based on the driver's response delay time.

[0052] On the other hand, the present invention provides a driving behavior monitoring storage medium based on a driving recorder, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the method described in any of the above-mentioned embodiments.

[0053] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0054] 1. This invention combines dashcam video data with vehicle status information to achieve high-precision tracking and correction of road participant trajectories. In particular, it introduces a curve-sensing trajectory correction method based on steering wheel angle, effectively eliminating trajectory distortion in the image coordinate system caused by vehicle turning and improving the accuracy of traffic event recognition. Based on this, an LSTM model is used for intelligent recognition of traffic motion events, combined with a driver response behavior judgment mechanism, enabling dynamic, real-time analysis and classification of driving behavior, thereby providing drivers with more accurate risk warnings and behavioral feedback.

[0055] 2. In this invention, a risk field strength model integrating physical dynamics and environmental factors is constructed, and a penalty mechanism for driver response delay is introduced to dynamically calculate the driving risk index. It not only considers key parameters such as relative speed and distance, but also combines external environmental factors such as road friction coefficient and regional risk level, which significantly improves the scientificity and adaptability of risk assessment. Attached Figure Description

[0056] Figure 1 This is a flowchart of the overall method of the present invention. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] Example 1: Please refer to Figure 1 As shown, this embodiment provides a driving behavior monitoring method based on a dashcam, including the following steps:

[0059] S1. Collect video data from the dashcam and extract image frames;

[0060] In this embodiment, during vehicle operation, a dashcam installed on the windshield, dashboard, or rearview mirror is used to collect real-time video streams of the road ahead. Subsequently, the continuous video stream is decoded frame by frame by the video decoding module, and key image frames are extracted according to a set time interval tn or frame rate to construct an image sequence arranged in chronological order.

[0061] S2. Perform multi-target detection on the image frame, identify the position information and category labels of road participants (vehicles and pedestrians) in the image frame, and construct the motion trajectory of road participants in the image sequence by combining the target tracking algorithm. At the same time, use the curve perception trajectory correction method based on steering wheel angle to perform rotation compensation and correction on the motion trajectory.

[0062] In this embodiment, multi-target detection is performed on the image frame to identify the location information and category labels of road participants (vehicles, pedestrians) in the image frame, including the following steps:

[0063] S2.1 Perform preprocessing operations on the image frames, including image scaling, normalization, and color space conversion, to ensure that the image meets the model input requirements;

[0064] S2.2. The preprocessed image frames are analyzed using the Multi-Object Detection Algorithm (YOLO) to identify road participants in the image. The basic components of the YOLO algorithm include an input processing module, a backbone feature extraction network, a multi-scale detection head, and an output decoding module. The input is a preprocessed (including scaling and normalization) fixed-size image (416×416 or 640×640). The backbone network (including CSPDarknet, Focus, and Backbone) is used to extract multi-layer semantic features of the image. The detection head (PANet or YOLO Head) predicts the position offset of candidate boxes, object class probabilities, and confidence scores at different scales. The YOLO model divides the image into several grids, each grid being responsible for predicting multiple object box information within its coverage area. The final output is a structured tensor containing all detected objects, including the bounding box coordinates, class label, and corresponding confidence score for each object.

[0065] S2.3 Output the bounding box coordinates, class probability distribution, and confidence score of each detected target road participant from the multi-object detection algorithm.

[0066] The method of constructing the motion trajectories of road participants in an image sequence using a target tracking algorithm includes the following steps:

[0067] S2.4 In each frame of the image, the identity management of each detected target (vehicle, pedestrian) output by the multi-target detection model is performed. If the target appears for the first time, a unique tracking identifier is assigned to it, and the corresponding tracker instance is initialized to record the target's state information and historical trajectory.

[0068] S2.5 For each detected target, crop an image patch within its bounding box area and input it into a deep neural network (ReID embedding model) to extract appearance feature vectors, which are used for appearance similarity judgment when associating across frames, to enhance tracking robustness, especially to maintain correct matching in scenarios where the target is partially occluded or temporarily lost.

[0069] S2.6. The target tracking algorithm (DeepSORT) is used to match the target detected in the current frame with the existing tracking trajectory in the previous frame. Combined with the target position predicted by the Kalman filter, the state of each tracked target is updated, including position, velocity and feature information, so as to achieve high-precision and continuous target tracking.

[0070] S2.7 Record the position information (i.e. bounding box coordinates) of each target with a tracking identifier in each frame of the image in chronological order, and construct a temporal trajectory data structure;

[0071] S2.8. Using a curve perception trajectory correction method based on steering wheel angle, perform rotation compensation and correction on the time-series trajectory data structure.

[0072] In existing solutions, target tracking (DeepSORT) primarily operates in the image coordinate system, recording the bounding box coordinates (pixel positions) of the target. However, when a vehicle turns, the dashcam's viewpoint (i.e., the image coordinate system) rotates relative to the world coordinate system. Stationary objects (including roadside trees and streetlights) will appear as arc-shaped trajectories in the image sequence, even though they are actually stationary in the real world. Similarly, the actual trajectories of other road users (vehicles and pedestrians) will also be distorted in the image coordinate system (i.e., a vehicle moving in a straight line may appear to be moving laterally in the image). The spatial relative relationships (vehicle distance change trend, relative speed direction) and speed change trends based on the distortion trajectory analysis become inaccurate (specifically, a vehicle traveling at a constant speed in the same direction as the dashcam on the outside of a curve will appear as a different object in the image trajectory). If the target appears to be rapidly approaching the vehicle laterally, this distorted trajectory, when input into the subsequent traffic event detection process (step S3), can lead to misjudgments (including false collision warnings and false entry behavior) or missed judgments (failure to identify genuine dangerous approach). The cornering perception trajectory correction method based on steering wheel angle utilizes readily available and high-precision steering wheel angle signals from the vehicle's CAN bus to estimate the vehicle's yaw rate (or approximately equivalent turning radius) in real time. It dynamically projects the trajectory points of the target in the image coordinate system back to the vehicle coordinate system, which is approximately aligned with the vehicle's forward direction, thus eliminating trajectory distortion caused by cornering (i.e., using the estimated yaw rate or turning radius to approximately compensate for the vehicle's rotational motion at each moment to eliminate trajectory distortion caused by cornering), resulting in a trajectory that better reflects the relative motion of the real physical world.

[0073] A curve perception trajectory correction method based on steering wheel angle is used to perform rotation compensation and correction on the time-series trajectory data structure, including the following steps:

[0074] S2.81. The vehicle's steering wheel angle signal is obtained in real time through the vehicle's OBD-II interface or by connecting to the CAN bus. The steering wheel angle signal represents the steering amplitude of the vehicle at time t, in degrees. Left turn is positive and right turn is negative. The steering wheel angle signal has the advantages of high precision, low latency and easy acquisition, providing input for subsequent trajectory rotation compensation.

[0075] S2.82. Based on the steering wheel angle and vehicle speed, calculate the vehicle's yaw rate to reflect the vehicle's current turning degree and rotation trend;

[0076] yaw rate for: ;

[0077] In the formula, The vehicle's longitudinal speed (in m / s, read from the vehicle's onboard diagnostic system). This is the actual steering angle of the front wheels (in rad), determined by the steering wheel angle. Transformed from ( In the formula, (Steering gear ratio) This refers to the vehicle's wheelbase (in meters, a fixed parameter for the vehicle). For the current frame;

[0078] S2.83. Extract the image coordinate information of each tracked target from the current image frame and its historical image frames (take the center point of the bounding box as the original position representation of the trajectory point);

[0079] S2.84. For each historical trajectory point, calculate the time difference between it and the current image frame, and combine it with the yaw rate at the corresponding moment to calculate the cumulative change in vehicle rotation angle within that time difference. , (In the formula, Indicates a historical frame. The time difference between frames (in seconds). , For historical frame indexing, including vehicles and pedestrians, among which, The sign is the same as the direction of the vehicle's rotation (left turn is positive, right turn is negative).

[0080] S2.85. Rotate the target point's coordinates in the image or camera coordinate system in the reverse direction by an angle to obtain the corrected trajectory coordinates. Based on the dashcam's camera calibration parameters (including focal length, principal point, and distortion coefficients), map the corrected trajectory coordinates to physical world coordinates in the vehicle coordinate system. (Specifically, using the pre-calibrated camera intrinsic matrix and distortion coefficients, perform distortion correction on the corrected trajectory points; based on the camera mounting height and extrinsic matrix, use inverse perspective transformation (IPM) to map the distortion-corrected image coordinates to physical world coordinates in the vehicle coordinate system.) Replace the historical trajectory coordinates of each target with the mapped trajectory coordinates to construct the final time-series trajectory data structure. ;

[0081] The target point's coordinates in the image or camera coordinate system are rotated in the opposite direction by the following angle:

[0082] ;

[0083] In the formula, The coordinates of the original trajectory points (in the image or camera coordinate system). These are the corrected trajectory coordinates.

[0084] S3. Analyze the spatial relative relationships and speed change trends among road participants based on motion trajectories, and use the LSTM model (Long Short-Term Memory Network Model) to identify traffic motion events.

[0085] In this embodiment, the spatial relative relationships and speed change trends among road participants are analyzed based on motion trajectories, and traffic motion events are identified using an LSTM model, including the following steps:

[0086] S3.1, For each pair of relevant objectives and (Including the relationship between the vehicle and the vehicle in front, and between the vehicle and pedestrians, where the vehicle is one's own vehicle), Euclidean distance and relative angle are calculated frame by frame based on the trajectory points;

[0087] Euclidean distance for:

[0088] ;

[0089] relative angle for:

[0090] ;

[0091] In the formula, , For target index;

[0092] S3.2, For each target It calculates the velocity vector at continuous time points, calculates the relative velocity vector (with the vehicle as a reference), and determines the direction and magnitude of the relative velocity (by combining the direction of the line connecting the two targets to determine whether it is a forward approach, a lateral crossing, or a moving away motion), which is used to analyze whether there are risky behavioral trends such as approaching, crossing, tailgating, or cutting in.

[0093] velocity vector for:

[0094] (Combined with the depth estimation module or known calibration information from within the camera, pixel velocity) Convert to actual speed (Unit: m / s, realizing velocity mapping from the image domain to the physical world coordinate system);

[0095] Relative velocity vector for:

[0096] ;

[0097] In the formula, For the goal The actual velocity vector;

[0098] S3.3 For each target or target pair (self vehicle and preceding vehicle), construct a trajectory-velocity-relative relationship sequence within a time window TN. In the formula, For the goal exist The velocity component in the lateral direction. For the goal exist The velocity component in the longitudinal direction, The spatial distance between the target and the vehicle. In this embodiment, relative speed (referring to the magnitude of relative speed) is used to ensure the calculated spatial distance between the target and the vehicle. Relative velocity To acquire physical meaning, the pixel-level trajectory point coordinates obtained in the image frame are converted into spatial coordinates in actual physical units. Specifically, this involves: using a monocular depth estimation method based on camera intrinsics or a depth sensing device installed on the vehicle (including binocular cameras and millimeter-wave radar) to obtain the target's depth in the current frame; performing back projection by combining the bounding box center point and camera calibration parameters (focal length, principal point, distortion coefficients, etc.); converting the image coordinates into physical world coordinates (using the method in step S2.85 for coordinate transformation); and then calculating the spatial distance between the target and the vehicle based on the physical world coordinates. (in meters) relative speed (in meters per second (m / s));

[0099] S3.4 Input the trajectory-velocity-relative relationship sequence into the pre-trained LSTM model and output the corresponding traffic event category prediction. The traffic event categories include sudden braking by the vehicle in front, sudden crossing by a pedestrian, and forced lane change by an oncoming vehicle. The pre-trained LSTM model takes the trajectory coordinates, velocity components, spatial distance, and relative speed of the target within the time window as temporal features as input. It extracts temporal dependency information through multiple stacked LSTM units and finally outputs the probability distribution of traffic event categories through fully connected layers and Softmax activation. The model is trained on labeled driving video trajectory data using cross-entropy loss to achieve accurate identification of traffic motion events such as sudden braking by the vehicle in front, pedestrian crossing, and forced lane change by an oncoming vehicle.

[0100] S4. Acquire and synchronize the vehicle's own state data (acceleration, steering angle, braking signal, etc.), and combine it with the identified traffic motion events to determine whether the driver has produced the corresponding driving behavior;

[0101] In this embodiment, the vehicle's own state data is acquired and synchronized, and combined with the identified traffic movement events, it is determined whether the driver has engaged in corresponding driving behavior, including the following steps:

[0102] S4.1 Acquire and synchronize the vehicle's own control status data, including steering wheel angle, brake pedal position, acceleration, vehicle speed, turn signal status, etc. Align with the identified traffic motion events using a unified timestamp, and define driver response behavior for each type of traffic event. If the traffic event is a sudden braking by the vehicle in front, the standard behavior response is an increase in braking signal and a negative jump in longitudinal acceleration (i.e., a significant decrease in acceleration). If the traffic event is a pedestrian suddenly crossing, the standard behavior response is deceleration (vehicle speed decreases) or steering to avoid the pedestrian (significant change in steering wheel angle). If the traffic event is a vehicle forcibly changing lanes, the standard behavior response is to maintain the lane or turn slightly to the outside (steering wheel angle deflection but no drastic lane change).

[0103] S4.2 Setting the response judgment time window The time interval (±1.5 seconds) is used as a reference, based on the central time point of traffic incident identification, to analyze the changes in vehicle control signals within that time interval; if in If braking, sudden steering, or sudden speed change is detected within seconds and the behavior matches the type of traffic incident, it is determined that the driver has made a reasonable response; if the change in control signal appears significantly delayed (more than 1 second) after the point of occurrence of the incident, it is determined to be a slow response; if there is no effective response action in the window, it is marked as no response; if the change in signal and the incident have a directional conflict (the driver continues to accelerate while facing a pedestrian crossing incident), it is determined to be dangerous behavior.

[0104] S4.3 Output the driver's response behavior judgment results to traffic events. The response behavior judgment results include the following types: normal response (timely and reasonable behavior), slow response (delayed response after the event occurs), no response (no corresponding action was collected), and dangerous behavior (reverse operation, such as running a red light, accelerating towards pedestrians, etc.).

[0105] S5. Establish a driving risk index to assess the risk of driving behavior, and make punitive adjustments to the driving risk index based on the driver's response delay time.

[0106] In this embodiment, a driving risk index is established to assess the risk of driving behavior, including the following steps:

[0107] S5.1 Extract risk factors from identified traffic incidents and driver response behaviors. Risk factors include the severity of the traffic incident (including sudden braking and sudden lane changes), the judgment result of the response behavior (normal response, slow response, no response, and dangerous behavior), and relative speed and distance (including the distance and speed difference between the driver and the vehicle or pedestrian in front).

[0108] S5.2 Normalize each risk factor to a uniform numerical range (0 to 1) to ensure comparability between different factors and facilitate weighted calculation;

[0109] S5.3 By constructing a risk field strength physical model, the risk field strength is calculated based on collision time, dynamic equivalent safe distance, and environmental and regional correction factors. The driver response delay is combined with a penalty correction to dynamically calculate the driving risk index. The risk field strength physical model is a modeling method that quantifies driving risk as physical field strength. It integrates dynamic parameters such as collision time (TTC), equivalent safe distance, environmental friction coefficient, and regional risk factors. By calculating the risk energy density at a specific moment, it reflects the strength of potential collision threats in the current traffic scenario. The driver response delay is combined with a penalty correction to finally output a comprehensive risk index that can change with time.

[0110] Traditional weighted summation methods cannot quantify the physical speed-distance coupling effect (the emergency risk of kinetic energy increasing with the square of speed in high-speed emergency braking scenarios), are difficult to dynamically respond to environmental changes (the decrease in friction coefficient due to rain or snow requires an equivalent shortening of the safety distance), and ignore the zero-tolerance risk characteristics of special areas (school zones). By constructing a risk field strength physical model, this paper innovatively integrates physical and environmental parameters such as collision time, dynamic friction coefficient, and regional multiplier, reconstructing the risk index into a field strength function based on vehicle dynamics principles. This fundamentally solves the problem of inaccurate modeling of static weights for scene complexity, environmental real-time performance, and regional specificities.

[0111] The specific steps involved in calculating the driving risk index using a physical model of risk field strength are as follows:

[0112] S5.31. Based on the vehicle dynamics model, combined with the vehicle's current speed, the relative speed between the vehicle and the target, the environmental friction coefficient, and the driver's reaction time, calculate the minimum safe braking distance under the current environment (by introducing the dynamic friction coefficient and the driver's reaction time, the abstract risk is transformed into a physically calculable minimum safe distance, solving the problem of underestimating the safe distance in the traditional model under low-adhesion road surfaces in rain and snow and fatigue driving scenarios). The vehicle dynamics model is based on the vehicle's structural parameters and current state. The inputs include information such as steering wheel angle, throttle / brake command, current vehicle speed, and road adhesion coefficient. By establishing the vehicle's motion equations in the longitudinal, lateral, and steering dimensions, the outputs the vehicle's motion state, such as position, speed, acceleration, heading angle, sideslip angle, and tire force, to simulate the vehicle's dynamic response under different driving operations and road conditions.

[0113] Minimum safe braking distance for:

[0114] ;

[0115] In the formula, This is the current road surface friction coefficient (dynamically acquired, adjusted for rain / snow). It is the acceleration due to gravity. For driver reaction time, For the vehicle's speed;

[0116] S5.32, Utilizing the spatial distance between the target and the vehicle And relative speed prediction of the time required for a vehicle to collide while maintaining its current operating state;

[0117] Time required for a collision for:

[0118] (when (Established at that time)

[0119] S5.33. Based on the spatial field modeling method, and by introducing the regional risk multiplier and collision time decay factor, calculate the risk field intensity at the current moment;

[0120] The risk field strength is:

[0121] ;

[0122] In the formula, The regional risk multiplier (used to increase the risk assessment weight in specific high-sensitivity areas (schools, hospitals); 2.0 for school / hospital areas and 1.0 for ordinary roads). This is the collision time decay factor (used to control the sensitivity of the risk field to time, default 2.0). The distance-to-sensitivity coefficient (a positive number) is calibrated using data.

[0123] S5.34. Combine the judgment results of the driver's response behavior to make penalty corrections. If the driver fails to respond to the traffic event normally or has a delayed response, a delay penalty factor is introduced to amplify the risk index and output the final driving risk index.

[0124] Final driving risk index for:

[0125] ;

[0126] In the formula, Driver response delay time (seconds) The delay penalty factor represents the weighting of delayed behavior on overall risk (0 for normal response, 1.0 for dangerous behavior), measured in s⁻. 1 ,s⁻ 1This indicates the percentage of risk penalty due to delay per second. If the driving behavior responds normally, then... The driving risk index remains unchanged;

[0127] S5.4. Based on the driving risk index, driving behavior is divided into multiple risk levels to facilitate decision-making and warning. The risk levels include low risk (driving risk index is less than the low risk threshold b), medium risk (driving risk index is greater than or equal to the low risk threshold b and less than the medium risk threshold c), high risk (driving risk index is greater than or equal to the medium risk threshold c and less than the high risk threshold d), and extremely high risk (driving risk index is greater than or equal to the high risk threshold d).

[0128] S5.5 Output risk assessment results, including driving risk index value, risk level label, and summaries of relevant traffic incidents and driving behaviors.

[0129] Example 2: This example provides a driving behavior monitoring device based on a dashcam, including:

[0130] The acquisition module is used to acquire video data from the dashcam in front of the vehicle and extract image frames;

[0131] The analysis module is used to perform multi-target detection on image frames, identify the location information and category labels of road participants in the image frames, and construct the motion trajectory of road participants in the image sequence by combining the target tracking algorithm. At the same time, the motion trajectory is rotated and corrected by using the curve perception trajectory correction method based on steering wheel angle.

[0132] The traffic incident detection module is used to analyze the spatial relative relationships and speed change trends among road participants, and to identify traffic motion events using a traffic incident detection model.

[0133] The driving behavior recognition module is used to acquire and synchronize the vehicle's own status data, and combine it with the identified traffic movement events to determine whether the driver has generated the corresponding driving behavior.

[0134] The risk assessment module is used to establish a driving risk index to assess the risk of driving behavior, and to make punitive adjustments to the driving risk index based on the driver's response delay time.

[0135] This embodiment also provides a driving behavior monitoring storage medium based on a dashcam, on which a computer program is stored, characterized in that: when the computer program is executed by a processor, it implements the steps of the method described in any of the above-mentioned embodiments.

[0136] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A driving behavior monitoring method based on a driving recorder, characterized in that, Includes the following steps: S1. Collect video data from the dashcam and extract image frames; S2. Perform multi-target detection on the image frame, identify the position information and category label of road participants in the image frame, and construct the motion trajectory of road participants in the image sequence by combining the target tracking algorithm. At the same time, use the curve perception trajectory correction method based on steering wheel angle to perform rotation compensation and correction on the motion trajectory. The method for correcting the curve perception trajectory based on steering wheel angle, which performs rotation compensation and correction on the motion trajectory, includes the following steps: The vehicle's steering wheel angle signal is obtained in real time through the vehicle's OBD-II interface. Based on the steering wheel angle and vehicle speed, the vehicle's yaw rate is calculated to reflect the vehicle's current degree of turning and rotation trend. Extract the image coordinate information of each tracked target from the current image frame and its historical image frames; For each historical trajectory point, calculate the time difference between it and the current image frame, and combine it with the yaw rate at the corresponding moment to calculate the cumulative change in the vehicle's rotation angle within that time difference; The target point's coordinates under the image are rotated in the opposite direction by an angle to obtain the corrected trajectory coordinates. Based on the dashcam's camera calibration parameters, the corrected trajectory coordinates are mapped to the physical coordinates under the vehicle coordinate system. The historical trajectory coordinates of each target are replaced with the mapped trajectory coordinates to construct the final time-series trajectory data structure. S3. Analyze the spatial relative relationships and speed change trends among road participants based on motion trajectories, and use the LSTM model to identify traffic motion events; S4. Acquire and synchronize the vehicle's own status data, and combine it with the identified traffic motion events to determine whether the driver has generated corresponding driving behavior; S5. Establish a driving risk index to assess the risk of driving behavior, and make punitive adjustments to the driving risk index based on the driver's response delay time. The establishment of a driving risk index to assess the risk of driving behavior includes the following steps: Extract risk factors from identified traffic incidents and driver response behaviors; Normalize each risk factor to a uniform numerical range; By constructing a physical model of risk field strength, the risk field strength is calculated based on collision time, dynamic equivalent safety distance, and environmental and regional correction factors. Penalty corrections are then applied in conjunction with driver response delay to dynamically calculate the driving risk index. Based on the driving risk index, driving behavior is divided into multiple risk levels; Output the risk assessment results.

2. The dashcam-based driving behavior monitoring method of claim 1, wherein: In step S2, multi-target detection is performed on the image frame to identify the location information and category labels of road participants in the image frame, including the following steps: S2.1 Perform preprocessing operations on the image frames; S2.2 Analyze the preprocessed image frames based on the multi-target detection algorithm to identify road participants in the image; S2.3 Output the bounding box coordinates, class probability distribution, and confidence score of each detected target road participant from the multi-object detection algorithm.

3. The driving behavior monitoring method based on a driving recorder according to claim 1, characterized in that: The method of constructing the motion trajectories of road participants in an image sequence using a target tracking algorithm includes the following steps: S2.4 In each frame of the image, identity management is performed on each detected target output by the multi-target detection model. If the target appears for the first time, a unique tracking identifier is assigned to it, and the corresponding tracker instance is initialized to record the target's state information and historical trajectory. S2.5 For each detected target, crop an image patch within its bounding box region and input it into a deep neural network to extract appearance feature vectors; S2.

6. Use the target tracking algorithm to match the target detected in the current frame with the existing tracking trajectory in the previous frame, and update the state of each tracked target by combining the target position predicted by the Kalman filter. S2.7 Record the position information of each target with a tracking identifier in each frame of the image in chronological order, and construct a temporal trajectory data structure; S2.

8. Using a curve perception trajectory correction method based on steering wheel angle, the time-series trajectory data structure is rotated and corrected.

4. The driving behavior monitoring method based on a driving recorder according to claim 1, characterized in that: In step S3, the spatial relative relationships and speed change trends among road participants are analyzed based on their motion trajectories, and traffic motion events are identified using an LSTM model. This includes the following steps: S3.1 For each pair of related targets, calculate the Euclidean distance and relative angle frame by frame based on their trajectory points; S3.2 For each target, calculate the velocity vector at consecutive time points, calculate the relative velocity vector, and determine the direction and magnitude of the relative velocity; S3.3 For each target, construct a trajectory-velocity-relative relationship sequence within a time window TN; S3.4 Input the trajectory-speed-relative relationship sequence into the pre-trained LSTM model and output the corresponding traffic event category prediction.

5. The driving behavior monitoring method based on a driving recorder according to claim 1, characterized in that: In step S4, the vehicle's own state data is acquired and synchronized, and combined with the identified traffic motion events, it is determined whether the driver has engaged in corresponding driving behavior, including the following steps: S4.1 Acquire and synchronize the vehicle's own control status data, align it with the identified traffic motion events through a unified timestamp, and define the driver's response behavior for each type of traffic event; S4.2 Set the response judgment time window ∆a seconds, and analyze the changes in vehicle control signals within this time interval based on the central time point of traffic event identification; S4.3 Output the judgment result of the driver's response behavior to traffic events.

6. The driving behavior monitoring method based on a driving recorder according to claim 1, characterized in that: The specific steps involved in calculating the driving risk index using a physical model of risk field strength are as follows: S5.

31. Based on the vehicle dynamics model, combined with the current speed of the vehicle, the relative speed between the vehicle and the target, the environmental friction coefficient and the driver's reaction time, calculate the minimum safe braking distance under the current environment. S5.

32. Predict the time required for a collision to occur while the vehicle is maintaining its current operating state by using the spatial distance and relative speed between the target and the vehicle. S5.

33. Based on the spatial field modeling method, and by introducing the regional risk multiplier and collision time decay factor, calculate the risk field intensity at the current moment; S5.

34. Adjust the penalty based on the driver's response behavior and output the final driving risk index.

7. A driving behavior monitoring device based on a dashcam, characterized in that, include: The acquisition module is used to acquire video data from the dashcam in front of the vehicle and extract image frames; The analysis module is used to perform multi-target detection on image frames, identify the location information and category labels of road participants in the image frames, and construct the motion trajectory of road participants in the image sequence by combining the target tracking algorithm. At the same time, the motion trajectory is rotated and corrected by using the curve perception trajectory correction method based on steering wheel angle. The method for correcting the trajectory of a vehicle by using a steering wheel angle-based curve perception includes the following steps: acquiring the vehicle's steering wheel angle signal in real time via the vehicle's OBD-II interface; calculating the vehicle's yaw rate based on the steering wheel angle and vehicle speed to reflect the vehicle's current turning degree and rotation trend; extracting image coordinate information of each tracked target from the current image frame and its historical image frames; calculating the time difference between each historical trajectory point and the current image frame, and calculating the cumulative rotation angle change of the vehicle within that time difference by combining the yaw rate at the corresponding moment; rotating the target point's coordinates in the image in the opposite direction by an angle to obtain the corrected trajectory coordinates; mapping the corrected trajectory coordinates to physical coordinates in the vehicle coordinate system based on the dashcam's camera calibration parameters; replacing the historical trajectory coordinates of each target with the mapped trajectory coordinates to construct the final time-series trajectory data structure. The traffic incident detection module is used to analyze the spatial relative relationships and speed change trends among road participants, and to identify traffic motion events using a traffic incident detection model. The driving behavior recognition module is used to acquire and synchronize the vehicle's own status data, and combine it with the identified traffic movement events to determine whether the driver has generated the corresponding driving behavior. The risk assessment module is used to establish a driving risk index to assess the risk of driving behavior, and to make punitive adjustments to the driving risk index based on the driver's response delay time. The process of establishing a driving risk index to assess the risk of driving behavior includes: extracting risk factors from identified traffic incidents and driver response behaviors; normalizing each risk factor to a uniform numerical range; constructing a physical model of the risk field strength, calculating the risk field strength based on collision time, dynamic equivalent safety distance, and environmental and regional correction factors, and applying punitive corrections based on driver response delays to dynamically calculate the driving risk index; classifying driving behavior into multiple risk levels based on the driving risk index; and outputting the risk assessment results.

8. A driving behavior monitoring storage medium based on a dashcam, wherein a computer program is stored thereon, characterized in that: When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images