Intelligent network connection vehicle risk avoiding method and system for longitudinal slope curve section of low-grade highway in mountainous area

Abstract

The invention relates to an intelligent network connection vehicle risk avoiding method and system for a mountain area low-grade road longitudinal slope curve road section. The method comprises the steps of collecting information of the mountain area road longitudinal slope curve road section; then fusing the various collected information by adopting a BEV Fusion model, and converting the information into unified BEV visual angle information; performing time sequence modeling and information completion by adopting a Transform model, constructing a unified BEV representation space, and obtaining all perception information; on the basis of perceived information, the possible obstacles on the road are comprehensively considered, multiple sub-scenes such as active avoidance, tailgating and interweaving are adopted for reverse driving between other participants and the CAV due to the narrow road, and the adaptability of the CAV in the complex traffic scene is improved; meanwhile, for possible interleaving conflicts, a dynamic interaction, prediction and safety risk avoiding decision comprehensive model is used, interaction and path decision optimization between the CAVs and other traffic participants are achieved, and the road passing efficiency and the safety level are improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent transportation technology in transportation engineering, and specifically relates to an intelligent connected vehicle hazard avoidance method and system for low-grade mountain roads with longitudinal slopes and curves. Background Technology

[0002] Connected Automatic Vehicles (CAVs) rapidly and accurately collect information through onboard and roadside facilities, employing relevant fusion algorithms to perceive the road environment in real time. They then interact and cooperate with other road users to control vehicle operation, largely solving the safety issues of human vehicles (HVs) on ordinary roads. However, for mountainous, low-grade highways with both CAVs and HVs, particularly on steep, winding sections, CAVs' visibility is limited by curves, slopes, and terrain. Furthermore, cost constraints limit the installation of roadside facilities on low-grade highways, restricting information perception. Although traffic volume is low, the presence of pedestrians, cyclists, and falling rocks remains. Combined with narrow roads, these factors exacerbate safety concerns. Therefore, for steep, winding sections of low-grade highways, leveraging the advantages of CAV technology to rapidly fuse and complete multi-source information to perceive all road conditions, and considering potential accident hazards and intersections in specific scenarios to implement avoidance measures and ensure traffic safety, has become one of the urgent technical challenges that CAV development needs to address.

[0003] Some existing research methods focus on hazard avoidance for HVs (Hard Vehicles), neglecting CAV (Cyclic Automotive Auxiliary Vehicle) hazard avoidance. These methods rely on driver visibility, experience, and front-end equipment, limiting information acquisition. While wide-angle mirrors (rearview mirrors or turning mirrors) and warning signs like speed reduction and VSM (Vehicle Management System) are used, drivers rely on horns and deceleration, lacking intelligent risk assessment, trajectory prediction, and safety decision-making based on smart perception, thus failing to meet the safe driving needs of CAVs. Other methods focus on conventional road scenarios, neglecting the unique and complex road sections like curves and slopes in low-grade mountainous areas, and cannot adapt to the information fusion and completion of CAV perception in obstructed environments. Still other methods fail to consider the mixed flow of CAVs, HVs, cyclists, and pedestrians on mountainous slopes and curves, as well as various possible complex scenarios, particularly situations involving falling rocks, oncoming traffic, and other road users not being able to see the CAVs.

[0004] Therefore, based on CAVs technology, traffic safety and control technology, and intelligent transportation systems, researching and inventing a CAVs safety avoidance method for mountainous low-grade highways with longitudinal slopes and curves is an urgent problem to be solved in the development of CAVs. This method involves "CAVs using onboard and roadside facilities to fuse and supplement information, comprehensively and three-dimensionally perceiving all information of the scene. If there are no obstacles or other traffic participants in the current lane, the vehicle can pass at a safe speed; if there are obstacles in the current lane that prevent passage and there are no participants in the opposite lane, the vehicle can use another lane, stop, or make a U-turn; if there are other traffic participants in both lanes and they cannot see the vehicle, the vehicle will take active deceleration, horn honking, or stop to avoid them; otherwise, the optimal driving path and driving strategy will be determined by using a collision risk field model, a participant trajectory prediction model, and a comprehensive safety decision model for low-grade mountain highways, and the vehicle can safely avoid danger by passing through the section." Summary of the Invention

[0005] To address the shortcomings of existing technologies, the technical problem this invention aims to solve is to provide an intelligent connected vehicle (CAV) hazard avoidance method and system for mountainous, low-grade highways with longitudinal slopes and curves. Based on the characteristics of CAVs and intelligent transportation, this method integrates and supplements various collected information, comprehensively perceives road environment information, and takes corresponding measures for various specific sub-scenarios to ensure traffic safety and reduce unnecessary accidents and potential property damage.

[0006] The solution adopted by the present invention to solve the aforementioned technical problem is:

[0007] In a first aspect, the present invention provides a method for intelligent connected vehicles to avoid risks on winding sections of low-grade mountain highways with longitudinal slopes, the method comprising the following steps:

[0008] Step 1: Collect information on the longitudinal slope and curved road sections of the mountain roads, including dynamic information collected by vehicle-mounted sensing devices, roadside equipment and V2X communication modules, as well as static basic information including road geometry;

[0009] The second step is to use the BEVFusion model to fuse various information collected, including road and terrain information, and convert it into unified BEV perspective information to obtain BEV feature maps.

[0010] Step 3: Combine the current frame BEV feature map output by the BEVFusion model with the historical BEV feature map to form a temporal feature sequence as input to the Transformer model. Use the Transformer model to perform temporal modeling and information completion, and output the completed BEV features to generate a complete representation of the dynamic state, including the current target type and target coordinate information, velocity, acceleration, and yaw angle, thus forming complete perception information.

[0011] Step 4: Determine the avoidance strategy based on whether the target type output by the Transformer model in Step 3 is a static obstacle or other traffic participants. If the Transformer model determines that the target type of the road segment ahead is a static obstacle (such as one caused by falling rocks), then determine the passage path based on the size and location of the static obstacle. If the Transformer model determines that the target type of the road segment ahead is other traffic participants, then a driving route and driving strategy need to be determined for avoidance driving.

[0012] The process of determining the driving route and driving strategy for risk avoidance includes: if the Transformer model determines that the target type output by the road segment ahead is other traffic participants, then the Transformer-based trajectory prediction model is used to preliminarily determine whether there are conflicting traffic participants in the lane where the CAV intends to drive, including those traveling in the opposite direction, crossing, or occupying the lane. If there are, and the conflicting traffic participants can see the vehicle, then the quantified safety risk characteristics are obtained through the mountain low-grade highway participant collision risk field model used to dynamically assess the potential collision risk of CAV in complex traffic environments. The quantified safety risk characteristics, the two-dimensional coordinate position and speed of the participants, and the relative distance and relative speed between the participants are input into the participant trajectory prediction model to predict the trajectory of the participants.

[0013] The pre-planned trajectory of this vehicle will be set. CAV (t+1) and the predicted trajectory Γ for each traffic participant j (t+1) Perform trajectory intersection detection to determine whether a conflict has occurred. If a conflict occurs, calculate the intersection angle α between the two trajectories to determine the severity of the conflict and prioritize them. Adopt different response strategies according to different priorities and make the safest and most appropriate driving decision in real time.

[0014] Furthermore, the collision risk field model for participants on low-grade mountain roads is derived from the lane risk field Risk. C Risk field with adjacent lanes A It consists of two parts. The current lane risk field is mainly used to assess the risks that the CAV may encounter in the current lane, specifically including the risk of rear-end collision caused by being too close to the vehicle in front, and the road boundary risk when approaching the right edge of the road (such as when there is no guardrail edge). The adjacent lane risk field is used to describe the risks faced by the CAV when traveling in another lane or changing lanes, mainly including the risk of head-on collision with oncoming traffic participants, and the left boundary risk when approaching the left edge. The ratio of the current lane risk to the adjacent lane risk is calculated, and the ratio of the risk values ​​is converted into a quantitative safety risk feature.

[0015] Furthermore, the risk field of this lane and the risk field of adjacent lanes are expressed by formulas (6) and (7), respectively:

[0016] Risk C =ω rear Risk rear +ω R Risk R (6)

[0017] Risk A =ω front Risk front +ω L Risk L (7)

[0018] Among them, Risk front With Risk rear These are the risk levels of a head-on collision and a rear-end collision for the participants, respectively. L With Risk R It is the risk intensity of the left and right boundary lines of the highway, ω front ω rear With ω L ω R These are their respective weighting coefficients, and the sum of the four is 1; among them,

[0019]

[0020] Among them, s r d r The longitudinal and lateral relative distances to the opposing participant are represented by n, where n is the number of participants; t represents time; and s represents the critical longitudinal distance s of the participants under the heading angle projection. c (t) and the distance d from the critical horizontal line c (t); longitudinal safety distance s s (t) and lateral safety distance d s (t); longitudinal risk component f s (s r ,t) and horizontal risk component f d (d r ,t);

[0021]

[0022] Among them, s' s (t) represents the longitudinal safety distance in a following scenario; W lane L is the lane width. p f' is the length of the participants. d (d r ,t) represents the lateral risk within the lane, f' s (s r ,t) represents the longitudinal risk component within the lane;

[0023]

[0024] Where θ is the polar angle of the CAV's location, and R(θ) is the polar radius of the vehicle's current position. L R R The left and right boundary lines are respectively, and k(θ) represents the road curvature.

[0025] Furthermore, the participant trajectory prediction model includes a Transformer module, a Graph Attention Network (GAT), and a fully connected layer. The temporal and interaction features of participants are encoded through the Transformer module and the GAT, respectively, capturing the evolutionary patterns of individual behavior and the dynamic interaction relationships among multiple participants. Finally, the outputs of the Transformer module and the GAT, along with high-dimensional quantified security risk features, are input into the fully connected layer and concatenated to obtain the output of the participant trajectory prediction model, used to predict the participant's trajectory at several future moments, as shown below:

[0026]

[0027] in, Let represent the predicted motion characteristics of participant j at time t+1, where These are the position, velocity, acceleration, and yaw angle of participant j at time t+1.

[0028] The quantitative safety risk features output by the participant collision risk field model are processed by a fully connected layer to obtain high-dimensional quantitative safety risk features.

[0029] Furthermore, different response strategies are adopted according to different priorities, specifically: when α∈[0°, 10°), it is a same-direction interference, the conflict level is low, and the response strategy is to maintain speed and observe; when α∈[10°, 30°), it is a vehicle turning ahead, the conflict level is medium, and the response strategy is to slow down and sound the horn, preparing to pass through the curve; when α∈[30°, 60°), it is a vehicle avoiding an obstacle ahead, the conflict level is medium, and the response strategy is to slow down and leave space for avoidance in advance; when α∈[60°, 120°), it is a vehicle crossing, the conflict level is medium, and the response strategy is to immediately slow down or brake and activate the warning lights; when α∈[120°, 150°), it is a vehicle using the wrong lane, the conflict level is high, and the response strategy is to actively slow down and avoid the obstacle and stop and wait; when α∈[150°, 180°], it is a vehicle driving in the wrong direction, the conflict level is high, and the response strategy is to actively slow down / wait, and stop and avoid the obstacle if necessary.

[0030] Furthermore, the specific process of determining the passage path based on the size and location of static obstacles is as follows:

[0031] a. If the obstacle is located in this lane and the width of the obstacle is greater than the width of this vehicle plus the lateral safety margin, it indicates that the vehicle cannot continue to move forward along this lane and turns to b; otherwise, continue to determine whether there are other traffic participants in this lane ahead. If not, determine the safe driving speed for this section of the road.

[0032] b. Determine if there are obstacles in the oncoming lane and their size. If there are obstacles and their width is greater than the width of the vehicle plus the lateral safety margin, it indicates that the road is blocked, and turn c. Otherwise, continue to determine if there are other road users in the oncoming lane. If not, determine the driving route and speed for using the other lane and proceed safely.

[0033] c. If there is not enough space, consider making a U-turn. Determine whether the vehicle can complete the U-turn based on its minimum turning radius. If the U-turn can be completed, then proceed with the decision to make the U-turn. Otherwise, the vehicle must be parked in a safe location and the hazard warning lights must be activated.

[0034] If other road users are present in the same or opposite lane, a driving route and strategy must be determined to avoid potential hazards.

[0035] Furthermore, during the process of determining the driving route and driving strategy for hazard avoidance, if conflicting road users cannot see the vehicle, the vehicle will take proactive measures to avoid them while determining the driving route and safe speed.

[0036] The calculation process for safe speed is as follows: First, use formula (1) to calculate the effective acceleration a after considering the slope, vehicle acceleration, and road conditions. eff Secondly, use formula (2) to calculate the safe distance D between vehicles. safe To avoid collisions; calculate the maximum safe speed v of the curve using formula (3). curve When a vehicle using another lane occupies the lane of the vehicle, in order to ensure safety, a sufficiently long avoidance distance needs to be determined so that the vehicle can stop or decelerate quickly. The avoidance speed v, which takes into account reaction time, is determined by formula (4). avoid Based on the above factors, the final safe speed v is determined as shown in formula (5). safe The calculations are shown below;

[0037] a eff =a acc -gsin(β)(1)

[0038]

[0039] Among them, a acc This is the acceleration of the vehicle on a curve with no longitudinal slope, where g is the acceleration due to gravity (9.8 m / s²). 2 ), β is the slope angle (positive for uphill, negative for downhill), v o This is the speed of the vehicle, v fWhere μ is the speed of other participants, R is the road friction coefficient, and D is the radius of curvature of the curve. brake T is the braking distance of the vehicle. reaction This is the reaction time, taken as 2.5s, v max This refers to the speed limit for mountain roads.

[0040] Secondly, the present invention provides an intelligent connected vehicle risk avoidance system for low-grade mountain highways with longitudinal slopes and curves. The system executes the method and includes roadside facilities, vehicle-mounted facilities, a GIS data acquisition unit, a V2X communication module, a participant collision risk field model, a participant trajectory prediction model, and a safety decision module.

[0041] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, can implement the steps of the method.

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

[0043] This invention integrates multi-source information from vehicle-mounted and roadside facilities to construct a unified BEV perspective, achieving perception information fusion, completion, and dynamic modeling. This significantly improves environmental understanding and decision-making robustness, reduces accident risk, and conducts risk assessment through a participant collision risk field model on low-grade mountain roads. A participant trajectory prediction model composed of Transformer and GAT models is used to predict participant trajectories, enabling real-time safe driving decisions and proactive risk avoidance control strategies when potential hazards arise. It is applicable to mountainous low-grade road sections with longitudinal slopes and curves, offering low cost, high perception accuracy, and strong coverage. It can be integrated as part of a CAV intelligent perception and autonomous driving system, improving safety levels and enhancing adaptability to this complex scenario.

[0044] (1) Unlike general information fusion and perception completion methods, this invention addresses the information limitation problem of longitudinal slope and curve sections on low-grade mountain roads. It fuses multi-source information collected from vehicle-mounted and roadside facilities, employs the BEVFusion model for multimodal data fusion, and the Transformer model for temporal modeling and information completion, constructing a unified BEV representation space to obtain comprehensive perception information and improve the understanding and decision robustness of CAVs in complex scenarios. Compared to methods that simply fuse camera information (Example: Method I), this method reduces the accident probability, average lateral and longitudinal root mean square acceleration, and average travel time by 77.3%, 30.2%, 33.3%, and 40.4%, respectively.

[0045] (2) Unlike general CAVs safe driving methods, the method of this invention aims to achieve safe driving on the longitudinal slope and curve sections of low-grade mountain roads. It comprehensively considers various sub-scenarios such as possible obstacles on the road, and the reverse driving between other participants and CAVs caused by narrow roads, and takes active avoidance, lane borrowing, and weaving, which increases the adaptability of CAVs in complex traffic scenarios.

[0046] (3) Unlike general CAVs safe driving methods, the method of this invention is specifically designed for the narrow road conditions of low-grade mountain roads with longitudinal slopes and curves. Based on all information obtained through intelligent perception, it utilizes a participant collision risk field model, a participant trajectory prediction model, and safety decision-making to optimize the interaction and path decision-making between CAVs and other traffic participants, thereby avoiding congestion and improving road traffic efficiency and safety. Compared to methods without risk assessment and interaction strategies (Example III), this method reduces the accident probability, root mean square of average lateral and longitudinal acceleration, and average travel time by 55.8%, 11.9%, 13.3%, and 24.4%, respectively. Attached Figure Description

[0047] Figure 1 This is a flowchart of a method for intelligent connected vehicles to perceive and avoid risks on a mountainous, low-grade highway with a longitudinal slope and curve, according to the present invention.

[0048] Figure 2 This is a flowchart of the BEVFusion model for multimodal data fusion in this invention;

[0049] Figure 3 This is a flowchart illustrating the target information completion process based on Transformer in this invention.

[0050] Figure 4 This is a schematic diagram of three driving routes for CAVs in a static obstacle sub-scenario of the present invention; wherein, (a) is the driving route when the CAV encounters an obstacle while driving in its own lane, but it does not affect the passage of other vehicles; (b) is the driving route when the CAV encounters an obstacle while driving in its own lane and cannot pass, and needs to use another lane (there are no other vehicles in the adjacent lane); (c) is the driving route when the CAV encounters an obstacle while driving in its own lane and there is no space for other vehicles to pass, and needs to make a U-turn.

[0051] Figure 5 This is a flowchart illustrating the CAV (Car Access Controller) process for determining driving lanes and driving strategies in this invention.

[0052] Figure 6 This is a schematic diagram illustrating whether a participant can see the CAV when it is obstructed by an obstacle in this invention;

[0053] Figure 7This is a flowchart of the integrated model for collision risk field, trajectory prediction, and safety decision-making of this CAV. Detailed Implementation

[0054] The present invention will be further explained below with reference to the embodiments and accompanying drawings, but this is not intended to limit the scope of protection of this application.

[0055] Example 1

[0056] This invention provides an intelligent connected vehicle (HV) hazard avoidance method for curved sections of low-grade mountain highways. Applicable to these sections, it enables environmental perception and hazard avoidance. The method targets curved sections of low-grade mountain highways, V2X (Vehicle-to-Everything) communication modules, wide-area radar and roadside equipment, vehicle-mounted vehicles (CAVs) and onboard sensing devices (onboard LiDAR, cameras, etc.), and potential cyclists, hazardous vehicles (HVs), and pedestrians. The specific steps are as follows:

[0057] Step 1: Collect information on the longitudinal slope and curved road sections of the mountain roads, including dynamic information collected by vehicle-mounted sensing devices, roadside equipment and V2X communication modules, as well as static basic information such as road geometry;

[0058] Step 2: Time synchronization of the collected multi-source data to ensure that images, point clouds, and terrain information are aligned under a unified timestamp; the BEVFusion model is used to fuse various collected information (including road and terrain information) and convert them into unified BEV perspective information. See below for the specific process. Figure 2 ;

[0059] ① Multimodal feature encoding: Leveraging the global context modeling capability of the Transformer network, image features F are extracted from multi-view RGB images via an image encoder. img The dense spatial structure of LiDAR point clouds is extracted using a sparse convolutional network to obtain LiDAR features F. lidar Simultaneously, static information such as road topography is combined with data transmitted by the V2X communication module, and terrain data features F are generated through perspective transformation. geo ;

[0060] ② Multimodal BEV feature construction: Constructing image features F img BEV image features are generated by projecting camera parameters onto the vehicle coordinate system. LiDAR feature F lidar Compressed mapping to BEVLiDAR features in the height dimension Terrain data features F geo A unified interpolation mapping is applied to the BEV plane and fused with BEV image features.

[0061] ③ Deep fusion of multimodal BEV features: The final BEV image features are fused with BEV LiDAR features, and a unified BEV feature map is generated by the BEV encoder. For details, see Liu et al.'s "Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation" (Liu Z, Tang H, Amini A, et al.Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation[C] / / 2023IEEE international conference on robotics and automation(ICRA).IEEE,2023:2774-2781.).

[0062] The obtained BEV feature maps can be used simultaneously for BEV map segmentation and 3D object detection tasks. BEV map segmentation can identify semantic regions such as lane lines, road boundaries, and passable areas; while 3D object detection can achieve three-dimensional spatial localization and category identification of traffic participants such as vehicles and pedestrians.

[0063] Step 3: Combine the current frame BEV feature map output by the BEVFusion model with historical BEV feature maps to form a temporal feature sequence, which is then used as input to the Transformer model. The Transformer model is used for temporal modeling and information completion, filling in perception gaps caused by occlusion or field-of-view limitations. The completed BEV features are output, generating a complete dynamic state representation including the current target type, target coordinates, velocity, acceleration, and yaw angle, forming complete perception information. Figure 3 As shown, the specific steps include:

[0064] ①BEVQueries Q Construction: Based on the feature map plane output by the BEVFusion model, a mesh-like learnable parameter Q is defined, where each Queries point represents the corresponding mesh cell region in the plane. The center of BEVQueries Q corresponds to the position of this CAV and can be customized according to road geometry characteristics (such as slope and curves).

[0065] ② Construction of historical frame feature sequences: Combine BEVQueries Q with historical BEV features B t-1 Together, they form a sequence, which serves as input for temporal modeling. Temporal features are extracted through a temporal self-attention mechanism, and then combined with spatial cross-attention to obtain multi-perspective spatial information.

[0066] ③ Temporal modeling module: Uses a Transformer-based temporal encoder to process historical BEV features B t-1 Modeling is performed using BEVQueries Q from B t-1 The system queries time information to capture target trajectories and environmental changes over time. Residual connections and normalization are applied between features to stabilize training, ultimately generating enhanced sequences that incorporate historical context.

[0067] ④ Spatial Information Completion: Utilizing a spatial cross-attention mechanism, specific methods can be found in the research on intelligent vehicle environmental perception technology based on multi-view fusion by He Hangtian (He Hangtian. Research on intelligent vehicle environmental perception technology based on multi-view fusion [D]. Jilin University, 2024). The output after temporal modeling is fused with the multimodal sensor information of the current frame to focus on missing region information. By projecting BEVQueries Q onto the image space, contextual information is extracted from multi-view images to improve the detection capability of occlusion and edge regions, and historical trajectory reasoning is combined to complete the perception of occlusion regions.

[0068] ⑤ Complete the BEV feature B t Output: The final output includes dynamic state information such as category, position, velocity, acceleration, and yaw angle, enabling comprehensive perception of both static and dynamic targets.

[0069] ⑥ Recursive update mechanism: After each frame of inference, the current B is updated. t Caching is used for modeling and fusion of the next frame, forming perceptual memory, which effectively improves the system's understanding and robustness to continuous dynamic scenes.

[0070] Step 4: Determine the target type based on the Transformer model from Step 3. Check if there are static obstacles ahead (such as those caused by falling rocks from the mountain). If there are, proceed to Step 5; otherwise, proceed to Step 6.

[0071] Step 5: Determine whether static obstacles ahead impede continued passage, and the appropriate route. See [link / reference needed]. Figure 4 As shown;

[0072] a. If a static obstacle is located in this lane, and the width of the static obstacle is greater than the width of this vehicle plus the lateral safety margin, it indicates that the vehicle cannot continue to move forward along this lane; proceed to step b. Otherwise, proceed to step six, decelerate, and proceed safely. See [link to step a] Figure 4 As shown in (a);

[0073] b. Determine if there are any static obstacles in the oncoming lane and their dimensions. If there are, and the obstacle's width is greater than the vehicle's width plus lateral safety margin, it indicates a road obstruction; proceed to step c. Otherwise, proceed to step seven to safely use the oncoming lane. Figure 4As shown in (b);

[0074] In this invention, "borrowing a lane" refers to the process by which a vehicle traveling in its lane, if its forward movement is obstructed, can pass through the obstructed section of road by directly borrowing a lane (when there are no vehicles traveling) or by exchanging lanes with vehicles in the opposite lane (when there are vehicles traveling). Low-grade highways and mountainous mountain slope curves are generally single-lane roads in one direction and do not allow driving in the opposite lane, but borrowing a lane is permitted when there are road obstacles.

[0075] c. If there is insufficient space, consider making a U-turn. Determine if this is feasible based on the vehicle's minimum turning radius. If a U-turn is possible and a decision is made, see [link to relevant documentation]. Figure 4 (c) indicates that you should proceed to step nine; otherwise, you must stop the vehicle in a safe location and activate the hazard warning lights.

[0076] Step 6: Based on the completed dynamic traffic participant coordinates, determine whether there are other traffic participants in the lane in this direction. If not, adjust the speed appropriately based on the speed limit and terrain features of the mountain road, and proceed to Step 9; otherwise, proceed to Step 8.

[0077] Step 7: Determine if there are other road users in the oncoming lane. If not, determine the route and speed for using the other lane, and proceed to Step 9; otherwise, proceed to Step 8 to determine the driving lane and formulate an appropriate driving strategy.

[0078] Step 8: Determine the driving lane and formulate an appropriate driving strategy. See the detailed process below. Figure 5 ;

[0079] ① The trajectory prediction model based on Transformer initially determines whether there are conflicting traffic participants such as those traveling in the opposite direction, crossing the road, or occupying the lane in the lane that this CAV intends to travel in. If not, the driving route and strategy are determined by "following the speed of the vehicle in front or adjusting the speed appropriately to drive safely", and then proceed to step nine; otherwise, proceed to step ②.

[0080] ② To determine whether participants in the conflict can see this vehicle, see... Figure 6 As shown, if not, turn to ③; otherwise, turn to ④;

[0081] ③ Determine the driving route and safe speed for "the vehicle to take measures such as actively slowing down, honking, and stopping to avoid participants." The calculation of the safe speed is as follows;

[0082] a eff =a acc -gsin(β)(1)

[0083]

[0084]

[0085] Among them, a acc This is the acceleration of the vehicle on a curve with no longitudinal slope, where g is the acceleration due to gravity (9.8 m / s²). 2 ), β is the slope angle (positive for uphill, negative for downhill), v o This is the speed of the vehicle, v f Where μ is the speed of other participants, R is the road friction coefficient, and D is the radius of curvature of the curve. brake T is the braking distance of the vehicle. reaction This is the reaction time, taken as 2.5 seconds; v max Speed ​​limits for mountain roads;

[0086] The calculation process for safe speed is as follows: First, use formula (1) to calculate the effective acceleration a after considering the slope, vehicle acceleration, and road conditions. eff Secondly, use formula (2) to calculate the safe distance D between vehicles. safe To avoid collisions (when v0 > v) f Effective immediately (ensuring the vehicle can decelerate to the speed of the vehicle in front); use formula (3) to calculate the maximum safe speed v of the curve. curve To prevent skidding and loss of control; when a vehicle using the other lane occupies the vehicle's lane or there is an obstacle, based on the available safe following distance D safe Braking distance D brake Calculate the maximum avoidance speed v avoid As shown in formula (4); ensure that the total stopping distance ≤ reaction distance + braking distance; taking into account all constraints, the final safe speed v is obtained by taking the minimum value through formula (5). safe To simultaneously meet the requirements of following vehicle safety, curve stability, obstacle avoidance feasibility, and legal speed limits;

[0087] ④ A comprehensive model consisting of a participant collision risk field model, a participant trajectory prediction model, and a safety decision module is used to determine the driving route and formulate a driving strategy. See the detailed process below. Figure 7 ,step:

[0088] a. Collect data, and obtain historical trajectory information of the target vehicle and surrounding participants through information fusion and completion. Set the scene where the target vehicle is located to have N surrounding targets, and the set of interactions between the target vehicle and the surrounding targets to be V. actor ={V1,V2,...,V N}, each target V i The motion state at time t is h i =[x i ,y i ,v i ,a i ,θ i ], where (x i ,y i ) represents two-dimensional coordinates, vi Indicates velocity, a i Represents acceleration, θ i Indicates the yaw angle;

[0089] b. Risk Assessment: Construct a participant collision risk field model for low-grade mountain roads. Using the participant characteristics (position, speed, etc.) and road linear characteristics (left and right boundary line positions, road curvature, etc.) collected above, dynamically assess the potential collision risk of CAVs in complex traffic environments. This model is based on the lane risk field. C Risk field with adjacent lanes A It consists of two parts, as shown in equations (6) and (7). The lane risk field is mainly used to assess the risks that the CAV may encounter in the current lane, specifically including the risk of rear-end collision caused by being too close to the vehicle in front, and the road boundary risk when approaching the right edge of the road (such as when there is no guardrail edge); the adjacent lane risk field is used to describe the risks faced by the CAV when traveling in another lane or changing lanes, mainly including the risk of head-on collision with oncoming traffic participants, and the left boundary risk when approaching the left edge.

[0090] Risk C =ω rear Risk rear +ω R Risk R (6)

[0091] Risk A =ω front Risk front +ω L Risk L (7)

[0092] Among them, Risk front With Risk rear These are the risk levels of a head-on collision and a rear-end collision for the participants, respectively. L With Risk R It is the risk intensity of the left and right boundary lines of the highway, ω front ω rear With ω L ω R These are their respective weighting coefficients.

[0093] In scenarios where vehicles share a road, the risk of a head-on collision with oncoming participants primarily stems from radial approach velocity. Both longitudinal and lateral displacements must be considered simultaneously. The risk intensity is derived as follows:

[0094]

[0095] Among them, L p With W pThe length and width of the participant, θ p (t) represents the heading angle of the participant at time t. as well as These are the velocity vector and position vector of the current CAV and participant j, respectively. CAV With W CAV The dimensions of this CAV vehicle are: length and width; s min and d min The minimum safe distances in the longitudinal and lateral directions are 3m and 1.75m, respectively. s,max a d,max These represent the maximum longitudinal deceleration and the maximum lateral deceleration of this CAV, respectively; s r With d r This represents the vertical and horizontal relative distance to the opposing participant, where n is the number of participants.

[0096] The calculation process for the risk intensity of a frontal collision is as follows: First, use formulas (8) and (9) to calculate the critical longitudinal distance s of the participant under the heading angle projection. c (t) and the distance d from the critical horizontal line c (t); then, the radial approach velocity Δv between the participants is calculated using formula (10). app (t); then use formulas (11) and (12) to calculate the longitudinal safety distance s respectively. s (t) and lateral safety distance d s (t); then, the longitudinal risk component f is calculated using formulas (13) and (14) respectively. s (s r ,t) and horizontal risk component f d (d r Finally, the maximum risk value is directly calculated using formula (15) as the risk intensity of a frontal collision. front .

[0097] The risk of a rear-end collision while following another vehicle primarily depends on the longitudinal relative speed, and its risk intensity is derived as follows:

[0098]

[0099] Among them, T h Indicates time interval, v C For the CAV vehicle speed, Δv s W represents the longitudinal relative velocity between this CAV and other participants. lane Indicates the width of the lane.

[0100] The calculation process for the risk intensity of a rear-end collision is as follows: First, the longitudinal safety distance s' in the following scenario is calculated using formula (16). s (t), and then use formula (17) to determine the lateral risk f' within the lane.d (d r Then, use formula (18) to calculate the longitudinal risk component f'. s (s r Finally, the maximum value of all risks is taken as the risk intensity (Risk) through formula (19). rear .

[0101] The risk intensity of the left and right boundary lines mainly depends on the location of the boundary lines and the road curvature. The derivation process is as follows:

[0102]

[0103] Where θ is the polar angle of the CAV's location, and R(θ) is the polar radius of the vehicle's current position. L R R The left and right boundary lines are respectively, and k(θ) represents the road curvature.

[0104] Weighting coefficient ω front ω rear With ω L ω R The determination process: Based on experience in perceiving potential risks on low-grade highways in mountainous areas, the weighting coefficient ω is determined. front ω rear With ω L ω R The weights are 0.4 and 0.2 respectively, and 0.2 and 0.2 respectively. Frontal collision risk is the main active risk faced by CAVs during driving. It has a high probability of occurrence and can be actively avoided through acceleration, deceleration and path planning. Therefore, it is given the highest weight of 0.4. Following collision risk describes the risk of rear-end collision caused by insufficient relative distance or excessive speed difference in following scenarios. Although its controllability is lower than that of forward collision risk, it has a significant impact in congested situations. The weight is set at 0.2. Left boundary risk and right boundary risk reflect the risk that CAVs may deviate from the road, collide with guardrails or enter dangerous areas when they approach the edge of the lane. This is especially critical on curves or narrow roads. Therefore, they are given a weight of 0.2 respectively.

[0105]

[0106] Formula (22) is used to calculate the ratio of the current lane risk to the adjacent lane risk. The risk value is then compared and converted into a quantitative safety risk characteristic, i.e., η, to determine the safety of passing through the adjacent lane. If the quantitative safety risk characteristic is greater than 1, it is beneficial for vehicles to pass through the adjacent lane; otherwise, it is necessary to maintain the original lane or wait.

[0107] c. The participant trajectory prediction model employs a structure that integrates a Transformer module and a Graph Attention Network (GAT) to jointly model different types of features. Specifically, the participant trajectory prediction model takes X as input. t As shown in Equation (23), it consists of three parts: the temporal features of the participants (including historical position and speed), the interaction features (including relative position and speed with other participants), and the quantified safety risk features (lane risk ratio). The temporal features and interaction features are encoded by the Transformer module and the Graph Attention Network (GAT), respectively, to capture the evolutionary patterns of individual behavior and the dynamic interaction relationships among multiple participants. Considering that the quantified safety risk features are usually low-dimensional scalar or vector information, a post-fusion strategy is adopted, using a fully connected layer (FC) to map the quantified safety risk features to the outputs [H1, H2, ... H] of the Transformer module and the Graph Attention Network (GAT). n ]、[I1,I2,…I n After matching the high-dimensional space, the three are then fused to obtain the output of the final participant trajectory prediction model.

[0108]

[0109] in Represents the two-dimensional coordinate position and velocity (temporal characteristics) of participant i. η represents the relative distance and relative speed (interaction feature) between participants. t This represents the lane risk ratio (quantifying safety risk characteristics).

[0110] Finally, the output of the participant trajectory prediction model is used to predict the participant's trajectory at several future moments, including information such as position, velocity, acceleration, and yaw angle, expressed as:

[0111]

[0112] in, Let represent the predicted motion characteristics of participant j at time t+1, where These are the position, velocity, acceleration, and yaw angle of participant j at time t+1.

[0113] d. Safety decision module, for the pre-planned trajectory of this vehicle. CAV (t+1) (set within the system) and the predicted trajectory Γ of each traffic participant j At (t+1), perform trajectory crossover detection to determine if a conflict has occurred.

[0114] ‖Γ CAV (t+1)-Γj (t+1)‖<d min_safe (25)

[0115] Where, d min_safe The minimum safe distance for a conflict is the Euclidean distance between the minimum longitudinal and lateral safe distances, taken as 3.47m.

[0116] If formula (25) is not valid, proceed to step nine; otherwise, it indicates that there is a potential conflict between this CAV and participant j. Calculate the intersection angle α between the two based on their trajectories to determine the severity of the conflict and prioritize them. Adopt different coping strategies according to different priorities as shown in Table 1, make the safest and most appropriate driving decision in real time, and proceed to step nine.

[0117] Table 1. Conflict Types, Levels, and Response Strategies

[0118]

[0119] Step 9: Follow the determined lane and driving strategy, and determine whether you have left this section of road. If not, turn back to step 1; otherwise, exit the intelligent perception and safety avoidance control system for this longitudinal slope curve.

[0120] In this invention, information such as GIS data of low-grade mountain roads, including road geometry, is provided by road authorities, traffic information companies, and CAVs information service providers. Data such as traffic flow status and other participants are collected in real time through roadside facilities and vehicle-mounted facilities.

[0121] In this invention, the vehicle's risk avoidance control strategy is automatically implemented by the perception and decision-making system of CAVs. Through the on-board control unit, combined with multimodal perception and V2X communication technology, when the vehicle enters the longitudinal slope and curve section of a low-grade mountain road, it realizes dynamic identification and avoidance behavior planning of forward targets, oncoming vehicles or potential risk areas.

[0122] The basic principle of this invention is as follows: various data collected from vehicle-mounted, roadside and other facilities are used to achieve semantic and geometric fusion of multi-source heterogeneous sensor data under a unified BEV perspective through the BEVFusion model. The Transformer model is used to complete information completion and realize perception of all information. Risk assessment is carried out through the participant collision risk field model of low-grade mountain roads. The participant trajectory prediction model composed of Transformer and GAT models is used to predict the trajectory of the participants, make safe driving decisions in real time, and take active risk avoidance control strategies when danger may occur.

[0123] Example 2

[0124] The CARLA 0.9.13 simulator was used as the simulation platform for the experiment (Tu Huizhao, Liu Jianquan, Wei Yutong, et al. Autonomous driving road test hazard scenario inference based on counterfactual inference [J]. Journal of Tongji University (Natural Science Edition), 2025, 53(02):223-232.). Multiple sensors (cameras, radar, and other vehicle-mounted and roadside facilities) in the CARLA simulator were used to simulate original elements such as vehicles on longitudinal slope and curved road sections in mountainous low-grade highways. Original data was collected for specific experimental scenarios, and multi-source information fusion and completion were performed. The vehicle was controlled to run and sample data on the Town 07 map via a Python interface. Table 2 shows the hardware and software configuration used in the simulation experiment. The parameters in the information fusion model, completion model, participant collision risk field model, and participant trajectory prediction model required for the experiment were collected based on samples from the simulation experiment and trained offline through a cloud platform.

[0125] The experimental road section was a two-lane, two-way road, 500m long, including a 300m curve; the longitudinal slope was 8%, the curve radius was 150m, and the speed limit was 30km / h. Other technical parameters, such as superelevation and lateral force coefficient, conformed to the design specifications for low-grade mountain roads. To eliminate unsteady-state influences and random factors, all tests were conducted 40 times, and the results were the average of the 40 tests. Each test lasted 11 hours, with the first hour discarded, meaning only 10 hours of data were collected for each test.

[0126] Table 2 Hardware and Software Configuration

[0127]

[0128] The experiment selected three types of indicators: safety, comfort, and efficiency. For details, please refer to the definitions and calculations in Shi Zixian's Improved Artificial Potential Field and Its Application in Autonomous Driving Decision-Making (Shi Zixian. Improved Artificial Potential Field and Its Application in Autonomous Driving Decision-Making [D]. South China University of Technology, 2022.), Tu Huizhao et al.'s Autonomous Driving Road Test Hazard Scenarios Based on Counterfactual Inference (Tu Huizhao, Liu Jianquan, Wei Yutong, et al. Autonomous Driving Road Test Hazard Scenarios Based on Counterfactual Inference [J]. Journal of Tongji University (Natural Science Edition), 2025, 53(02):223-232.), and Qin Yanyan et al.'s Basic Traffic Flow Map and Driving Comfort of Autonomous Driving Vehicles (Qin Yanyan, Wang Hao. Basic Traffic Flow Map and Driving Comfort of Autonomous Driving Vehicles [C] / Proceedings of the 13th China Intelligent Transportation Conference, 2018.).

[0129] ① Safety indicator selection: accident probability C p It compares the headway TH with the collision headway threshold TH. safe (Select 1.55s) is obtained by the following formula;

[0130] Cp =N C / N(26)

[0131] Where N C The total number of accidents, i.e., the headway TH is less than the collision headway threshold TH. safe The total number of collisions includes collisions with HVs, cyclists, pedestrians, and possible obstacles such as mountains, trees, and rolling stones; N is the total number of vehicles.

[0132] The relationship between the accident probability and the actual accident rate in traffic engineering statistics can be found in the conversion formula between the accident rate and the accident probability calculated by Zhang Shan et al. (Zhang Shan, Pang Mingbao, Ren Boning. Microscopic simulation and accident rate study of traffic flow under icy and snowy road conditions [J]. China Safety Science Journal, 2020, 30(1):148-154.).

[0133] ② Comfort index selection: The root mean square of lateral acceleration and the root mean square of longitudinal acceleration of the vehicle are used as evaluation indicators for lateral comfort and longitudinal comfort, respectively.

[0134] ③ The efficiency index is represented by the average travel time T of each vehicle through this road segment.

[0135] To verify the effectiveness of the "Intelligent Connected Vehicle Risk Avoidance Method for Longitudinal Slope and Curve Sections of Low-Grade Highways in Mountainous Areas" (Method IV) proposed in this invention, three other typical methods were constructed for comparative experiments.

[0136] Method 1: This is a traditional single-modal method that uses images acquired by an onboard camera as the sole sensing source. It does not introduce information fusion and completion mechanisms, and the vehicle relies solely on image recognition results for braking control. See Wang et al.'s method (Wang Y, Guizilini V C, Zhang T, et al. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries[C] / / Conference on Robot Learning.PMLR,2022:180-191.)

[0137] Method II: The method of constructing BEV representation using LSS (Lift-Splat-Shoot) generates single-frame BEV map features by fusing on-board and road test data. It has certain spatial modeling capabilities, but lacks temporal modeling and perceptual completion in dynamic scenes; see Philion et al.'s method (Philion J, Fidler S. Lift, Splat, Shoot: Encoding Images From Arbitrary Camera Rigs by Implicitly Unprojecting to 3D[C] / / ComputerVision-ECCV 2020:16th EuropeanConference, Glasgow, UK, August 23-28, 2020:194-210.).

[0138] Method III: This method employs intelligent perception using BEVFusion and Transformer, but without risk assessment or interaction strategies. In other words, this method relies solely on multimodal fusion-based perception information to determine potential traffic conflicts and adopts a relatively conservative "passive parking and waiting" strategy without risk assessment or interaction.

[0139] Table 3 shows the results of comparing the four methods using evaluation indicators. Experimental results indicate that:

[0140] Table 3 Comparison of performance indicators of different methods

[0141]

[0142] (1) Method I (traditional single-modal method) suffers from incomplete information in perception and delays in decision-making, which leads to a higher probability of collision; the root mean square of lateral and longitudinal acceleration is large, resulting in poor comfort and a long average driving time. This is because there may be dangerous and unseen scenarios on this road section, making driving behavior dangerous and unstable. Once such scenarios occur, such as not being able to see other participants, emergency deceleration will result in poor comfort, slow decision-making, and a long average driving time.

[0143] (2) Method II (the method of generating BEV representation by LSS) reduces the collision probability and improves safety by introducing BEV; the root mean square of lateral and longitudinal acceleration is slightly lower, reflecting the advantage of this method in perception accuracy. However, due to the lack of temporal modeling, the overall driving comfort is insufficient; the lack of dynamic interaction results in a slightly longer driving time.

[0144] (3) Using method III (the method without risk assessment and interaction strategy), the safety is better, which is due to the adoption of a more conservative strategy; at the same time, the acceleration is relatively stable; the driving time will be slightly longer due to the conservative strategy of not actively intervening.

[0145] (4) Method IV (the intelligent connected vehicle avoidance method for low-grade mountain roads with longitudinal slopes and curves in this invention) has the best safety, fast and intelligent response; the root mean square of acceleration is the lowest, enabling efficient perception and decision-making, and the driving process is smooth and comfortable; its efficient perception and decision-making capabilities avoid unnecessary delays and excessive braking, and the overall driving time is the shortest.

[0146] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0147] Any aspects not covered in this invention are applicable to existing technologies.

Claims

1. A method for intelligent connected vehicles to avoid hazards on winding sections of low-grade mountain highways with longitudinal slopes, characterized in that, The risk avoidance method Includes the following steps: Step 1: Collect information on the longitudinal slope and curved road sections of the mountain roads, including dynamic information collected by vehicle-mounted sensing devices, roadside equipment and V2X communication modules, as well as static basic information including road geometry; The second step is to use the BEVFusion model to fuse various information collected, including road and terrain information, and convert it into unified BEV perspective information to obtain BEV feature maps. Step 3: Combine the current frame BEV feature map output by the BEVFusion model with the historical BEV feature map to form a temporal feature sequence as input to the Transformer model. Use the Transformer model to perform temporal modeling and information completion, and output the completed BEV features to generate a complete representation of the dynamic state, including the current target type and target coordinate information, velocity, acceleration, and yaw angle, thus forming complete perception information. Step 4: Determine the avoidance strategy based on whether the target type output by the Transformer model in Step 3 is a static obstacle or other traffic participants. If the Transformer model determines that the target type of the road segment ahead is a static obstacle (such as one caused by falling rocks), then determine the passage path based on the size and location of the static obstacle. If the Transformer model determines that the target type of the road segment ahead is other traffic participants, then a driving route and driving strategy need to be determined for avoidance driving. The process of determining the driving route and driving strategy for risk avoidance includes: if the Transformer model determines that the target type output by the road segment ahead is other traffic participants, then the Transformer-based trajectory prediction model is used to preliminarily determine whether there are conflicting traffic participants in the lane where the CAV intends to drive, including those traveling in the opposite direction, crossing, or occupying the lane. If there are, and the conflicting traffic participants can see the vehicle, then the quantified safety risk characteristics are obtained through the mountain low-grade highway participant collision risk field model used to dynamically assess the potential collision risk of CAV in complex traffic environments. The quantified safety risk characteristics, the two-dimensional coordinate position and speed of the participants, and the relative distance and relative speed between the participants are input into the participant trajectory prediction model to predict the trajectory of the participants. The pre-planned trajectory of this vehicle will be set. CAV (t+1) and the predicted trajectory Γ for each traffic participant j (t+1) Perform trajectory intersection detection to determine whether a conflict has occurred. If a conflict occurs, calculate the intersection angle α between the two trajectories to determine the severity of the conflict and prioritize them. Adopt different response strategies according to different priorities and make the safest and most appropriate driving decision in real time.

2. The risk avoidance method according to claim 1, characterized in that, The participant collision risk field model for low-grade mountain roads is derived from the lane risk field Risk. C Risk field with adjacent lanes A It consists of two parts. The current lane risk field is mainly used to assess the risks that the CAV may encounter in the current lane, specifically including the risk of rear-end collision caused by being too close to the vehicle in front, and the road boundary risk when approaching the right edge of the road (such as when there is no guardrail edge). The adjacent lane risk field is used to describe the risks faced by the CAV when traveling in another lane or changing lanes, mainly including the risk of head-on collision with oncoming traffic participants, and the left boundary risk when approaching the left edge. The ratio of the current lane risk to the adjacent lane risk is calculated, and the ratio of the risk values ​​is converted into a quantitative safety risk feature.

3. The risk avoidance method according to claim 1, characterized in that, The risk field of this lane and the risk field of the adjacent lane are expressed by formulas (6) and (7), respectively: Risk C =ω rear Risk rear +ω R Risk R (6) Risk A =ω front Risk front +ω L Risk L (7) Among them, Risk front With Risk rear These are the risk levels of a head-on collision and a rear-end collision for the participants, respectively. L With Risk R It is the risk intensity of the left and right boundary lines of the highway, ω front ω rear With ω L ω R These are their respective weighting coefficients, and the sum of the four is 1; among them, Among them, s r d r The longitudinal and lateral relative distances to the opposing participant are represented by n, where n is the number of participants; t represents time; and s represents the critical longitudinal distance s of the participants under the heading angle projection. c (t) and the distance d from the critical horizontal line c (t); longitudinal safety distance s s (t) and lateral safety distance d s (t); longitudinal risk component f s (s r ,t) and horizontal risk component f d (d r ,t); Among them, s' s (t) represents the longitudinal safety distance in a following scenario; W lane L is the lane width. p f' is the length of the participants. d (d r ,t) represents the lateral risk within the lane, f' s (s r ,t) represents the longitudinal risk component within the lane; Where θ is the polar angle of the CAV's location, and R(θ) is the polar radius of the vehicle's current position. L R R The left and right boundary lines are respectively, and k(θ) represents the road curvature.

4. The risk avoidance method according to claim 1, characterized in that, The participant trajectory prediction model includes a Transformer module, a Graph Attention Network (GAT), and a fully connected layer. The temporal and interaction features of participants are encoded by the Transformer module and the GAT, respectively, capturing the evolutionary patterns of individual behavior and the dynamic interaction relationships among multiple participants. Finally, the outputs of the Transformer module and the GAT, along with high-dimensional quantified security risk features, are input into the fully connected layer and concatenated to obtain the output of the participant trajectory prediction model, used to predict the participant's trajectory at several future moments, as shown below: in, Let represent the predicted motion characteristics of participant j at time t+1, where These are the position, velocity, acceleration, and yaw angle of participant j at time t+1.

5. The risk avoidance method according to claim 1, characterized in that, The different response strategies adopted according to different priorities are as follows: When α∈[0°, 10°), it is a same-direction interference, the conflict level is low, and the response strategy is to maintain speed and observe; when α∈[10°, 30°), it is a vehicle turning ahead, the conflict level is medium, and the response strategy is to slow down and sound the horn to prepare to pass through the curve; when α∈[30°, 60°), it is a vehicle avoiding an obstacle ahead, the conflict level is medium, and the response strategy is to slow down and leave room for avoidance in advance; when α∈[60°, 120°), it is a vehicle crossing over, the conflict level is medium, and the response strategy is to immediately slow down or brake and activate the warning lights; when α∈[120°, 150°), it is a vehicle using the wrong lane, the conflict level is high, and the response strategy is to actively slow down and avoid the obstacle and stop and wait; when α∈[150°, 180°], it is a vehicle driving in the wrong direction, the conflict level is high, and the response strategy is to actively slow down / wait, and stop and avoid the obstacle if necessary.

6. The risk avoidance method according to claim 1, characterized in that, The specific process of determining the passage path based on the size and location of static obstacles is as follows: a. If the obstacle is located in this lane and the width of the obstacle is greater than the width of this vehicle plus the lateral safety margin, it indicates that the vehicle cannot continue to move forward along this lane and turns to b; otherwise, continue to determine whether there are other traffic participants in this lane ahead. If not, determine the safe driving speed for this section of the road. b. Determine if there are obstacles in the oncoming lane and their size. If there are obstacles and their width is greater than the width of the vehicle plus the lateral safety margin, it indicates that the road is blocked, and turn c. Otherwise, continue to determine if there are other road users in the oncoming lane. If not, determine the driving route and speed for using the other lane and proceed safely. c. If there is not enough space, consider making a U-turn. Determine whether the vehicle can complete the U-turn based on its minimum turning radius. If the U-turn can be completed, then proceed with the decision to make the U-turn. Otherwise, the vehicle must be parked in a safe location and the hazard warning lights must be activated. If other road users are present in the same or opposite lane, a driving route and strategy must be determined to avoid potential hazards.

7. The risk avoidance method according to claim 1, characterized in that, During the process of determining a driving route and strategy for hazard avoidance, if conflicting road users cannot see the vehicle, the vehicle will take proactive measures to avoid them, determine the driving route and safe speed. The calculation process for safe speed is as follows: First, use formula (1) to calculate the effective acceleration a after considering the slope, vehicle acceleration, and road conditions. eff Secondly, use formula (2) to calculate the safe distance D between vehicles. safe To avoid collisions; Calculate the maximum safe speed v of the curve using formula (3). curve When a vehicle using another lane occupies the lane of the vehicle, in order to ensure safety, a sufficiently long avoidance distance needs to be determined so that the vehicle can stop or decelerate quickly. The avoidance speed v, which takes into account reaction time, is determined by formula (4). avoid Based on the above factors, the final safe speed v is determined as shown in formula (5). safe The calculations are shown below; a eff =a acc -gsin(β) (1) Among them, a acc This is the acceleration of the vehicle on a curve with no longitudinal slope, where g is the acceleration due to gravity (9.8 m / s²). 2 ), β is the slope angle (positive for uphill, negative for downhill), v o This is the speed of the vehicle, v f Where μ is the speed of other participants, R is the road friction coefficient, and D is the radius of curvature of the curve. brake T is the braking distance of the vehicle. reaction This is the reaction time, taken as 2.5s, v max This refers to the speed limit for mountain roads.

8. A smart connected vehicle hazard avoidance system for low-grade mountain highways with longitudinal slopes and curves, characterized in that, The system executes the method of any one of claims 1-7, including roadside facilities, vehicle-mounted facilities, GIS data acquisition unit, V2X communication module, participant collision risk field model, participant trajectory prediction model, and safety decision module.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program can perform the steps of the method described in any one of claims 1-7.

Images