A traffic operation risk dynamic evaluation method based on multi-source data fusion

Abstract

This invention discloses a dynamic assessment method for traffic operation risks based on multi-source data fusion, belonging to the field of traffic operation risk technology. It includes real-time acquisition of multi-source traffic data from vehicle-mounted sensors, roadside units, meteorological data platforms, and high-precision maps; a dynamic parameter engine to generate an adaptive repulsive field model; a risk occlusion compensation model to calculate the coupled risk field strength of multiple obstacles; and another risk occlusion compensation model to calculate the coupled risk field strength of multiple obstacles. This invention overcomes the perception limitations of traditional single sensors through multi-level data collaboration among vehicle-mounted sensors, roadside V2X units, meteorological data, and high-precision maps. It achieves early detection of obstacles in blind spots and comprehensive characterization of the road environment, solving the perception lag problem in extreme scenarios. Based on dynamic parameters such as road curvature and road surface friction coefficient, it adjusts the potential energy field gain in real time, enabling the model to adapt to the risk characteristics of different road conditions and avoiding risk misjudgment by traditional fixed-parameter models.

Description

Technical Field

[0001] This invention relates to the field of traffic operation risk technology, specifically to a dynamic assessment method for traffic operation risks based on multi-source data fusion. Background Technology

[0002] Vehicle operation risk assessment is a core component of Intelligent Transportation Systems (ITS) and autonomous driving technologies, providing safety warnings for vehicle decision-making by quantifying potential risks in the traffic environment. Early assessment methods primarily relied on single-sensor data (such as radar ranging) or static models (such as safe distance models), which suffer from insufficient real-time performance when dealing with complex scenarios such as multi-vehicle interactions and sudden weather changes. With the development of perception technologies, risk modeling methods based on potential field theory have achieved risk visualization by simulating physical field effects; however, existing technologies still face significant bottlenecks in dynamic adaptability, multi-source data fusion, and handling complex scenarios.

[0003] Traditional methods rely solely on onboard sensors (such as LiDAR and cameras) to acquire localized data, lacking collaboration with roadside V2X units, meteorological platforms, and high-precision maps. This results in delayed perception of blind-spot obstacles (such as vehicles cutting in from the side or rear), especially in extreme weather conditions like rain and fog, where the detection range and accuracy of a single sensor decrease significantly, failing to comprehensively characterize the road environment. For example, the dynamic risk potential energy field model proposed in Chinese patent CN112002143B is based solely on onboard sensor data without incorporating external environmental data. In scenarios with visibility below 100m, the response delay to blind-spot targets exceeds 200ms.

[0004] The gain parameters (such as the repulsive field coefficient) of existing potential field models are usually preset to fixed values, without considering the influence of dynamic environmental factors such as road friction coefficient and road curvature on the risk field. For example, in scenarios with slippery roads (friction coefficient < 0.5) or small-radius curves (curvature radius R < 200m), fixed parameter models cannot adaptively adjust risk sensitivity, resulting in delayed or misjudged deceleration suggestions, increasing the risk of collision.

[0005] The current model uses a linear superposition method to calculate the risk field strength of multiple obstacles, without considering the influence of obstacle spatial overlap and relative motion direction. When scenarios such as large vehicles obstructing pedestrians or multi-vehicle pileups occur, it cannot accurately quantify the nonlinear transmission characteristics of risk. For example, the existing technology lacks an occlusion compensation mechanism based on the Rotated Intersection over Union (IOU), and when the obstacle spatial overlap IOU > 0.6, the risk field strength calculation error exceeds 30%.

[0006] Existing methods lack efficient multi-source data temporal alignment mechanisms, suffer from insufficient Doppler motion compensation in LiDAR point clouds, and exhibit significant timestamp discrepancies (typically >100ms) between V2X and sensor data, leading to data fusion distortion. Furthermore, inadequate edge computing architecture deployment places an excessive burden on onboard computing power, making it difficult to complete complex risk field calculations within 80ms and failing to meet the real-time requirements of autonomous driving.

[0007] Based on this, the present invention designs a dynamic assessment method for traffic operation risks based on multi-source data fusion to solve the above problems. Summary of the Invention

[0008] To address the aforementioned shortcomings of existing technologies, this invention provides a method for dynamic assessment of traffic operation risks based on multi-source data fusion.

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

[0010] A dynamic assessment method for traffic operation risks based on multi-source data fusion includes the following steps:

[0011] S1: Real-time acquisition of multi-source traffic data from vehicle sensors, roadside units, meteorological data platforms, and high-precision maps, including vehicle motion status, blind spot obstacle vectors, visibility, road surface friction coefficient, and road curvature;

[0012] S2: Through the dynamic parameter engine, the potential energy field gain parameters are dynamically adjusted according to visibility, road surface friction coefficient and road curvature to generate an adaptive repulsive force field model.

[0013] S3: Based on the risk occlusion compensation model, the spatial distribution of obstacles and their relative motion direction are integrated to calculate the risk field strength of multiple obstacles.

[0014] Preferably, the dynamic parameter engine uses a Bayesian network to dynamically adjust the weights of multi-source data. When there is a conflict between vehicle-mounted and roadside data, the weight of the roadside data is increased to above 0.7 using the posterior probability P(weight|observation).

[0015] Preferably, the dynamic adjustment of the potential energy field gain parameter in S2 includes the road repulsion field gain coefficient λ according to the formula. Calculate, where R is the radius of curvature, k = 0.01m. 1 The curvature response coefficient is calibrated by fitting the actual vehicle sideslip angle, and μ is the road friction coefficient compensation factor.

[0016] The velocity repulsion field gain coefficient is introduced into the dynamic parameter α optimized by federated learning, and the velocity repulsion field model is updated using the measured value m from the on-board mass sensor.

[0017] Preferably, the implementation of the risk occlusion compensation model in S3 includes calculating the spatial overlap of obstacles based on the Rotated IOU, and triggering compensation when IOU > 0.6;

[0018] The deformable attention mechanism of the DETR model is used to identify occluded regions and output an occlusion relationship probability matrix. When the occlusion probability is ≥85%, the occlusion coefficient correction is activated.

[0019] Preferably, multi-source data time-series alignment includes:

[0020] Doppler motion compensation is performed on the lidar point cloud, and the point cloud position offset Δd = V d ×Δt, where ;

[0021] The road surface friction coefficient μ is updated by fusing data from the onboard IMU and roadside humidity sensor using Kalman filtering. The update formula is as follows: .

[0022] Preferably, a single lane is divided into 20cm×20cm grid units, and a local risk increment matrix is ​​generated based on the detection of road surface anomalies by the roadside radar-visual integrated machine. ;

[0023] By using a genetic algorithm to optimize the scheduling of edge computing nodes, the collaborative latency between roadside video stream processing and vehicle-mounted risk field calculation is controlled within 80ms.

[0024] Solve the transformation matrix .

[0025] Preferably, the calibration of roadside equipment installation deviations includes:

[0026] Extracting lane line intersections from high-precision maps ;

[0027] Matching the detection point set of the Rave integrated machine ;

[0028] The transformation matrix T is solved by using a quaternion rotation matrix, so that the grid positioning error is ≤5cm.

[0029] Preferably, the weight allocation of the dynamic parameter engine includes:

[0030] When visibility is less than 100m, the DETR model is compressed to 12M parameters through knowledge distillation, reducing the LiDAR weight to 0.3 and increasing the V2X communication weight to 0.7.

[0031] When the road surface moisture content is greater than 0.5, the friction coefficient compensation factor... By combining Shannon's information entropy theory, the uncertainty of environmental perception can be reduced.

[0032] Preferably, the edge computing architecture deployment includes:

[0033] The roadside edge nodes perform point cloud distortion compensation and V2X data alignment, and the on-board computing unit calculates the dynamic risk potential energy field in real time.

[0034] The output layer generates machine-readable codes containing risk level, grid value, and waypoint quadruples, where the risk level is stored using a 1-5 level enumeration type.

[0035] Preferably, the risk enhancement in the curve region includes calculating the centrifugal potential energy. ;

[0036] Normalized to risk value ;

[0037] When the radius of curvature R < 200m, the curvature response coefficient k = 0.01m is fused through a Bayesian network. -1 ,Will This is superimposed on the overall risk field.

[0038] Compared with the prior art, the beneficial effects of this invention are as follows:

[0039] 1. This invention overcomes the perception limitations of traditional single sensors by using multi-level data collaboration among vehicle-mounted sensors, roadside V2X units, meteorological data, and high-precision maps. It enables early detection of obstacles in blind spots and comprehensive characterization of the road environment, solving the problem of perception lag in extreme scenarios. Based on dynamic parameters such as road curvature and road surface friction coefficient, the potential energy field gain is adjusted in real time, enabling the model to adapt to the risk characteristics of different road conditions, avoiding risk misjudgment by traditional fixed parameter models, and improving the accuracy of assessment in complex scenarios.

[0040] 2. This invention constructs a risk coupling model with a nonlinear pattern based on the cross-union ratio and the dynamic occlusion coefficient of the motion direction, accurately quantifies the risk transmission characteristics in multi-obstacle occlusion scenarios, solves the evaluation distortion problem of traditional linear superposition models, and ensures cross-source data synchronization through a multi-source data time-series alignment mechanism (including Doppler motion compensation and timestamp interpolation), reduces the impact of data latency on risk assessment, and improves the real-time performance of the system.

[0041] 3. The dynamic parameter engine of this invention adaptively adjusts the weights of multi-source data according to meteorological conditions (such as visibility and road surface humidity), optimizes the sensor fusion strategy in adverse environments such as rain, fog, and slippery conditions, enhances the robustness of the system, and adopts an edge computing architecture to deploy data processing and risk calculation tasks separately, reducing the on-board computing power burden. At the same time, it achieves efficient transmission and application of assessment results through machine-readable risk coding.

[0042] 4. This invention introduces a centrifugal potential energy enhancement mechanism for curved areas, and combines it with dynamic correction of risk field strength based on curvature radius to improve the risk assessment system under special road conditions and enhance the comprehensiveness of driving safety warnings. The risk occlusion compensation model effectively addresses complex scenarios such as large vehicles obstructing pedestrians and multi-vehicle chain interactions by integrating the spatial overlap of obstacles and the relative motion direction, making the risk field calculation more consistent with actual traffic patterns. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0044] Figure 1 This is a three-dimensional distribution map of the curve obstruction risk field in a traffic operation risk dynamic assessment method based on multi-source data fusion according to the present invention. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0046] Example 1: A method for handling curve occlusion based on Bayesian networks and DETR;

[0047] Through multi-source data fusion, the vehicle-mounted LiDAR (128 lines) data is processed by cubic spline interpolation algorithm to achieve spatiotemporal alignment and ensure that the error is within ±50ms. The roadside RSU millimeter-wave radar data is processed by Laplace smoothing algorithm to handle missing values ​​and improve data integrity. The high-precision map provides a curve curvature radius of 50m, and the meteorological data records a humidity of 65% and a corresponding friction coefficient of 0.4.

[0048] A 6-node Bayesian network model was constructed using conflict resolution methods. The nodes cover vehicle confidence, roadside data integrity, and meteorological data reliability. The posterior probabilities of the Bayesian network were then utilized. The fusion weights of multi-source data are dynamically adjusted. When there is a conflict between vehicle-mounted and roadside data, the weight allocation is optimized based on the Bayesian network calculation results. For example, the weight of roadside data is increased to 0.7 to ensure the reliability of the fused data.

[0049] Using an occlusion inference method, the image from the vehicle-mounted camera (1280×720 resolution) is input into the DETR model of a 3-layer Transformer encoder. This model focuses on the occlusion region through a deformable attention mechanism and outputs an occlusion probability matrix. When the occlusion probability of the construction vehicle reaches 92%, the rotation intersection-over-union (R-IOU) calculation is initiated, with the formula as follows:

[0050]

[0051] Where θ is the angle between the relative motion directions, and the measured value is 30°. This calculation can quantify the spatial relationship between the shading area and the vehicle's driving path.

[0052] The curvature compensation coefficient, obtained by fitting 1000 sets of curve sideslip angle data using the risk field calculation method, is k=0.01m. -1 Goodness of fit R 2 =0.92. Introducing the dynamic parameter α=0.1kg / m into the velocity repulsion field formula, the calculated repulsion field energy E4=1200N·m. Combined with the collision time TTC=1.8s, and based on the risk field calculation results, a braking suggestion is generated, recommending a deceleration of 3.2m / s². 2 .

[0053] In occluded scenarios, the risk identification accuracy is improved to 92%, which is a significant improvement compared to the 78% of traditional methods. The spatiotemporal alignment delay is controlled within 50ms, meeting the real-time requirements.

[0054] like Figure 1 As shown, the risk field strength in the curve region exhibits a distribution characteristic of being high on the inner side and low on the outer side, consistent with the logic of centrifugal potential energy superposition. The V2X weight of 0.7 and the lidar weight of 0.33 in the figure correspond to the sensor fusion strategy under rain and fog conditions. The peak risk field strength in the right blind zone is 2.5N. m) The obstacle repulsion field dominated by roadside V2X data verifies the compensation effect of multi-source data fusion on blind spot risk. In addition, the smooth gradient distribution of risk field strength confirms the accuracy control of grid positioning error ≤5cm.

[0055] Example 2: This example further discloses a method for treating slippery road surfaces based on Kalman filtering and federated learning;

[0056] The road surface friction coefficient μ is updated by fusing onboard IMU acceleration data (accuracy ±0.1 m / s²) with roadside surface humidity sensor data using Kalman filtering. The update formula is as follows:

[0057]

[0058] The final calculated road surface friction coefficient μ=0.3. A DETR model was deployed at the roadside node (Jetson Orin) to process the real-time video stream (30fps), while a Bayesian network was run on the in-vehicle unit (EyeQ6) for data fusion. A genetic algorithm was used for scheduling, reducing processing latency to below 80ms. A federated learning framework was employed to perform data anonymization on the local device, removing vehicle IDs and driver features to comply with Article 30 of the GDPR.

[0059] By employing the FedAvg algorithm to aggregate parameters across regions, the model's convergence speed was improved by 37%. Introducing Δa = -1.2 m / s² into the velocity repulsion field formula, combined with the curvature compensation parameter λ = 1.5, a speed limit recommendation of 30 km / h was calculated. Verification based on Shannon's information entropy theory showed that multi-source data fusion reduced information entropy by 30%, significantly decreasing the uncertainty in environmental perception.

[0060] Under extreme weather conditions, the perception latency is controlled within 100ms, and the energy consumption of edge computing is reduced by 40% compared to a single node.

[0061] Example 3: This example further discloses a multi-source data collaborative intersection decision-making method;

[0062] A hierarchical attention mechanism is used to fuse data from vehicle-mounted cameras (1920×1080 resolution), roadside V2X units (communication latency less than 10ms), high-precision maps (using OpenDrive format), and weather stations (wind speed 5m / s) across modalities, generating feature vectors that include traffic light status and pedestrian trajectories. A Bayesian network is used to dynamically adjust data weights, for example, vehicle-mounted confidence is 0.8 and roadside integrity is 0.9. When the DETR model identifies an 85% probability of pedestrian occlusion at the intersection, a rotation IOU and TTC-weighted occlusion compensation mechanism is triggered, with the following formula:

[0063]

[0064] When TTC < 2s and R-IOU > 0.6, video stream processing at the roadside node has a computational capacity of 15 TOPS, while risk field calculation at the on-board unit has a computational capacity of 5 TOPS. A 20ms cycle timer is achieved through a priority queue, and optimization using a genetic algorithm reduces the average system latency to 75ms. Using this method, the accuracy of intersection decision-making is improved to 95%.

[0065] Example 4: This example further discloses a multi-source data compensation method in a tunnel environment;

[0066] In tunnel environments, cubic spline interpolation is used to perform spatiotemporal alignment of vehicle-mounted LiDAR data, with the error controlled within ±50ms. Simultaneously, Kalman filtering is employed to process data transmitted from the roadside RSU to estimate vehicle position, achieving an accuracy of ±0.5m.

[0067] Inside the tunnel, consider a radius of curvature R = 200m and a curvature response coefficient k = 0.01m. -1 The method was determined through theoretical derivation combined with simulation verification. The derivation is based on the vehicle dynamics formula k=V² / (gR) and verified through a simulation environment. ΔV=5m / s was introduced into the velocity repulsion field formula, resulting in E4=800N·m. Considering the collision time TTC=2.5s, the recommended safe following distance is 30m. The vehicle coordinate system was converted to a high-precision map coordinate system using a quaternion rotation matrix, with the conversion error controlled to less than 0.1°. A federated learning framework was employed to ensure local data processing, meeting the requirements of Article 21 of the Cybersecurity Law of the People's Republic of China. Using this method, the positioning error in tunnel scenarios is less than 1 meter, and the delay in risk field calculation does not exceed 90ms.

[0068] Example 5: Robust optimization method for heavy rain scenarios;

[0069] The vehicle-mounted camera utilizes Laplace smoothing technology to handle rain and fog noise, while the roadside millimeter-wave radar predicts vehicle trajectories using Kalman filtering, with the prediction error controlled within ±0.3m. Simultaneously, meteorological data (50mm / h rainfall) is updated locally using a federated learning framework, optimizing the road surface friction coefficient μ to 0.25. A Bayesian network incorporating meteorological reliability nodes is established, based on posterior probabilities. The weight of roadside sensors is increased to 0.8. When there is a conflict between vehicle-mounted and roadside data, roadside data (such as vehicle location information) is prioritized for decision-making.

[0070] Knowledge distillation technology was used to compress the DETR model to a 12M parameter scale, significantly improving inference speed by 3 times (processing time reduced from 120ms to 40ms) while keeping accuracy loss to less than 2%. Roadside nodes (Jetson Orin) handled video stream processing, and the vehicle-mounted unit handled risk field calculations. The overall processing latency was controlled within 100ms. The velocity repulsion field formula incorporated Δa = -1.5m / s², combined with a curvature compensation coefficient λ = 1.8, to calculate a deceleration suggestion with a target speed of 20km / h. Verified by Shannon information entropy theory, the information entropy was reduced by 30% after multi-source data fusion, effectively improving the reliability of environmental perception. In rainy weather scenarios, the perception error was controlled to less than 10%.

[0071] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for dynamic assessment of traffic operation risks based on multi-source data fusion, characterized in that, Includes the following steps: S1: Real-time acquisition of multi-source traffic data from vehicle sensors, roadside units, meteorological data platforms, and high-precision maps, including vehicle motion status, blind spot obstacle vectors, visibility, road surface friction coefficient, and road curvature; S2: Through the dynamic parameter engine, the potential energy field gain parameters are dynamically adjusted according to visibility, road surface friction coefficient and road curvature to generate an adaptive repulsive force field model. Dynamically adjusting the potential energy field gain parameters, including the road repulsion field gain coefficient λ, according to the formula... Calculate, where R is the radius of curvature, k = 0.01m. 1 The curvature response coefficient is calibrated by fitting the actual vehicle sideslip angle, and μ is the road friction coefficient compensation factor. The velocity repulsion field gain coefficient is introduced into the dynamic parameter α optimized by federated learning, and the velocity repulsion field model is updated by the measured value m of the on-board mass sensor. S3: Based on the risk occlusion compensation model, the spatial distribution of obstacles and the relative motion direction are integrated to calculate the risk field strength of multiple obstacles coupled together; The implementation of the risk occlusion compensation model includes calculating the spatial overlap of obstacles based on the intersection-union ratio of a rotated rectangle, and triggering compensation when IOU > 0.6; The deformable attention mechanism of the DETR model is used to identify occluded regions and output an occlusion relationship probability matrix. When the occlusion probability is ≥85%, the occlusion coefficient correction is activated. The single lane is divided into 20cm×20cm grid units. Based on the roadside radar-visual integrated machine detecting road surface anomalies, a local risk increment matrix is ​​generated. ; A genetic algorithm is used to optimize the scheduling of edge computing nodes, controlling the collaborative latency between roadside video stream processing and vehicle-mounted risk field calculation to within 80ms; the transformation matrix is ​​solved. ; Roadside equipment installation deviation calibration includes: extracting the lane line intersection set from the high-precision map. Matching the detection point set of the Ravec all-in-one machine The transformation matrix T is solved by using a quaternion rotation matrix, so that the grid positioning error is ≤5cm. Increased risk in curve regions, including calculation of centrifugal potential energy. Normalized to risk value When the radius of curvature R < 200m, the curvature response coefficient k = 0.01m is fused through a Bayesian network. -1 ,Will This is superimposed on the overall risk field.

2. The method for dynamic assessment of traffic operation risk based on multi-source data fusion according to claim 1, characterized in that, The dynamic parameter engine uses a Bayesian network to dynamically adjust the weights of multi-source data. When there is a conflict between vehicle-mounted and roadside data, the weight of the roadside data is increased to above 0.7 by using the posterior probability P(weight|observation).

3. The method for dynamic assessment of traffic operation risk based on multi-source data fusion according to claim 1, characterized in that, Multi-source data time series alignment includes: Doppler motion compensation is performed on the lidar point cloud, and the point cloud position offset Δd = V d ×Δt, where ; The road surface friction coefficient μ is updated by fusing data from the onboard IMU and roadside humidity sensor using Kalman filtering. The update formula is as follows: .

4. The method for dynamic assessment of traffic operation risk based on multi-source data fusion according to claim 1, characterized in that, The weight allocation of the dynamic parameter engine includes: When visibility is less than 100m, the DETR model is compressed to 12M parameters through knowledge distillation, reducing the LiDAR weight to 0.3 and increasing the V2X communication weight to 0.

7. When the road surface moisture content is greater than 0.5, the friction coefficient compensation factor... By combining Shannon's information entropy theory, the uncertainty of environmental perception can be reduced.

5. The method for dynamic assessment of traffic operation risk based on multi-source data fusion according to claim 1, characterized in that, Edge computing architecture deployment includes: Roadside edge nodes perform point cloud distortion compensation and V2X data alignment, and the on-board computing unit calculates the dynamic risk potential energy field in real time. The output layer generates machine-readable codes containing risk level, grid value, and waypoint quadruples, where the risk level is stored using a 1-5 level enumeration type.

Images