A risk assessment and security constraint generation method, device, equipment and medium

Abstract

The present application belongs to the technical field of automatic driving, and particularly relates to a risk assessment and safety constraint generation method, device, equipment and medium, the method comprising: acquiring perception information and converting it into a structured scene state; combining road topology and lane connection relationship to predict the trajectories of surrounding traffic participants, and using lateral offset and heading deviation to correct online rematching, to obtain predicted trajectories conforming to road constraints; establishing a potential collision damage model based on motion state and category, and mapping it into a current real-time risk value through a time-space decay function; determining the interaction type according to the spatial relationship between the maximum risk object and the ego vehicle, and generating longitudinal, lateral and / or lane-changing safety rules; and fusing the output safety constraint set for the planning layer and the control layer, including planning constraints, cost penalties or control triggers.

Description

Technical Field

[0001] This invention belongs to the field of autonomous driving technology, specifically relating to a risk assessment and safety constraint generation method, apparatus, device, and medium. Background Technology

[0002] As the application of autonomous vehicles gradually expands from closed parks and highways to urban roads, the vehicle operating environment exhibits significant characteristics such as complex road structures, diverse types of traffic participants, strong interactive relationships, and high uncertainty in future behavior. In urban road scenarios, autonomous vehicles not only need to meet road geometric constraints, traffic rule constraints, and their own dynamic constraints, but also must engage in continuous behavioral games and dynamic interactions with surrounding traffic participants such as motor vehicles, non-motor vehicles, and pedestrians. Therefore, the system safety requirements are significantly higher than in traditional driving scenarios.

[0003] Currently, mainstream trajectory planning methods mostly rely on cost function optimization, obstacle avoidance distance constraints, or rule-based behavioral decision-making mechanisms to jointly generate paths and speeds. These methods can output feasible trajectories that meet basic safety requirements in conventional traffic scenarios. However, in highly interactive and uncertain scenarios (such as intersections without traffic lights, roundabouts, and congested mixed-traffic areas), relying solely on preset safety distances, collision detection thresholds, or penalty terms in the cost function often fails to characterize the potential conflict risks between vehicles and surrounding traffic participants in real time and continuously. Specifically, static or fixed-threshold safety distances cannot reflect dynamically changing levels of danger, while discrete collision detection easily overlooks the process of risk accumulation and evolution within the prediction time domain.

[0004] On the other hand, traditional Time to Collision (TTC) and its various variants (such as improved TTC and reciprocal time-distance) are mostly based on longitudinal relative motion models, primarily applicable to following or rear-end collision scenarios on highways. For the lateral and longitudinal coupling conflicts prevalent in urban roads—such as left-turning vehicles and straight-going non-motorized vehicles at intersections, lane-changing vehicles and parallel vehicles in adjacent lanes, pedestrians and turning vehicles—these one-dimensional relative motion-based indicators are prone to numerical instability, unclear physical meaning, or misjudgment of safe / dangerous states when quantifying conflict risk. Therefore, existing TTC-type indicators are insufficient to effectively support the risk assessment needs in complex urban environments.

[0005] In summary, there is an urgent need to propose a risk assessment method that can comprehensively consider road topology constraints, differences in traffic participant categories, predicted potential collision consequences in the time domain, and spatiotemporal decay relationships. This method should be able to directly transform the risk assessment results into constraints and safety control triggering criteria in the trajectory planning process, thereby improving the safe driving capability of autonomous vehicles in urban road scenarios with strong interaction and high uncertainty. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, the present invention provides a risk assessment and safety constraint generation method, apparatus, device and medium, which aims to improve the safe driving capability of autonomous vehicles in urban road scenarios with strong interaction and strong uncertainty.

[0007] A first aspect of this invention provides a risk assessment and security constraint generation method, the method comprising: S1. Obtain perception information of the environment in which the autonomous vehicle is located, and convert the perception information into structured scene state information; S2. Based on the structured scene state information, combined with the road topology and lane connection relationship, predict the future trajectory of each of the surrounding traffic participants in the prediction time domain, and in the prediction process, perform online rematching correction on the prediction results according to the lateral offset and heading deviation of the surrounding traffic participants relative to the matched lanes, so as to obtain the predicted trajectory that meets the road constraints. S3. Within the prediction time domain, based on the motion state parameters of the vehicle and each of the surrounding traffic participants and the category information of the traffic participants, a potential collision damage model is established to obtain the potential collision damage value of the vehicle and each of the surrounding traffic participants at each prediction time, and the potential collision damage value is mapped to the real-time risk value at the current time through a spatiotemporal decay function. S4. Determine the maximum risk object based on the real-time risk value, and determine the interaction type based on the spatial relationship between the vehicle and the maximum risk object. Generate a safety rule including at least one of longitudinal safety rules, lateral safety rules and lane-changing safety rules based on the interaction type. S5. The safety rules are fused to output a set of safety constraints for the planning layer and the control layer. The output form of the safety constraint set includes at least one of the following: planning constraints for screening candidate trajectories, cost penalties for increasing the priority of safe trajectories in the cost function, and control triggers for triggering active safety actions.

[0008] Preferably, the step of predicting the future trajectories of each of the surrounding traffic participants in the prediction time domain includes: S201. For each of the surrounding traffic participants, based on their current motion state, a coarse predicted trajectory is generated by iteratively extrapolating using a constant rotation rate and constant acceleration model within the total predicted time, according to a preset departure step length, and the iterative extrapolation satisfies the following formula: in,( ) for the surrounding traffic participants in the first The predicted location coordinates at each time point , , , The surrounding traffic participants are respectively in the first The velocity, heading angle, longitudinal acceleration, and yaw rate at each predicted moment, where Δt is the distance from the walk; Lane matching is performed based on the coarse predicted trajectory, and the matching methods include: S202. When the surrounding traffic participants first enter the prediction loop, global lane matching is performed on the pre-constructed map lane centerline point set to obtain the currently matched lane. S203. When the surrounding traffic participants have been continuously tracked, based on the lane matching result of the previous moment and the lane topology connection relationship, local lane matching is performed within the candidate range formed by the previous matched lane and its subsequent connected lanes to obtain the current matched lane.

[0009] Preferably, the process of obtaining the predicted trajectory that conforms to the road constraints based on the currently matched lane includes: S2011. Calculate the distance cost and heading cost between the end point of the coarse predicted trajectory and the candidate lane, respectively, and construct a comprehensive cost function: in, Candidate lanes The overall cost, , These are the distance weighting coefficient and the heading weighting coefficient, respectively. The Euclidean distance between two points. The endpoint of the coarsely predicted trajectory. Candidate lanes The center point, For the coarse trajectory end heading, To match lane heading; S2012. Select the candidate lane that corresponds to the minimum value of the comprehensive cost as the target lane; S2013. When the length of a single target lane is insufficient to support the complete predicted distance, the target lane and its subsequent connecting lanes are centerline-stitched according to the lane topology connection relationship to form a continuous reference path. : In this context, ⊕ indicates that the centerline is spliced ​​according to the lane connection relationship. To meet the predicted length requirements for the number of subsequent lane layers, To match lane line information in the lane.

[0010] Preferably, obtaining the predicted trajectory conforming to road constraints based on the continuous reference path includes: S2014. Taking the current position of the surrounding traffic participants as the starting point and the nearest matching point of the predicted trajectory length on the continuous reference path as the ending point, a cubic Bézier curve is used to correct the predicted trajectory to obtain the predicted trajectory that conforms to the road constraints. The cubic Bézier curve satisfies: in, The current location of the surrounding traffic participants. , Control points of the Bézier curve To predict trajectory length In the continuous reference path the nearest matching point on, , These are the tangent vectors at the starting and ending points, respectively; S2015. During the prediction process, the lateral offset and heading consistency of the surrounding traffic participants relative to the currently matched lane are monitored in real time. When one of the following conditions is met, the re-matching of the currently matched lane and the regeneration of the predicted trajectory are triggered: in, The width of the currently matched lane. The heading of the currently matched lane. Provides real-time location and heading information for surrounding traffic participants.

[0011] Preferably, the step of establishing a potential collision damage model to obtain the potential collision damage values ​​between the vehicle and each of the surrounding traffic participants at each predicted time includes: S301, within the prediction time domain, for the first... The surrounding traffic participants, in the first The motion state parameters of the vehicle and the surrounding traffic participants are obtained at each predicted time point. These motion state parameters include the vehicle's motion state parameters at the predicted time point. The speed at each predicted time point , No. The surrounding traffic participants in the first The speed at each predicted time point And the two in the Relative velocity at each predicted time point ; S302. Based on the motion state parameters, construct the comprehensive velocity characterization term according to the following formula. : in, The speed difference weighting coefficient; S303. Perform a nonlinear mapping on the comprehensive velocity characterization term to obtain the basic damage quantity. : in, , These are nonlinear mapping parameters.

[0012] Preferably, mapping the potential collision damage value to a real-time risk value at the current moment using a spatiotemporal decay function includes: S304, Based on the traffic participant category information, the first... The weighting coefficients of the matching categories of the surrounding traffic participants Construct the first formula as follows The surrounding traffic participants in the first Potential collision damage value at each predicted time : S305. Constructing time decay and spatial decay functions using exponential functions: in, , Control the rate of decay of time risk and the rate of decay of spatial risk respectively. Let these be the spatial decay function and the temporal decay function. For the current moment, For the first The surrounding traffic participants in the first Distance from the predicted time position to the predicted trajectory endpoint; The difference between the current time and the predicted terminal time. S306. The real-time risk value at the current moment is calculated based on the potential collision damage value, the time decay function, and the spatial decay function.

[0013] Preferably, the interaction type includes at least one of longitudinal interaction within the same lane, lateral interaction between adjacent lanes, and lane-changing interaction; When the interaction type is longitudinal interaction within the same lane, the minimum longitudinal safety distance is calculated based on the worst-case assumption using the following formula. The longitudinal safety rule is generated based on the minimum longitudinal safety distance: in, For reaction time, For maximum acceleration, For minimum braking deceleration, For maximum braking deceleration, , These are the speeds of the following vehicle and the vehicle in front, respectively. , These are the lengths of the rear vehicle and the front vehicle, respectively. When the interaction type is lateral interaction between adjacent lanes, in the Frenet coordinate system or the lane local coordinate system, based on the worst-case lateral approach process where the vehicle and the maximum risk object move towards each other with maximum lateral acceleration during the reaction time, and both decelerate with minimum lateral braking acceleration after the reaction time ends until their lateral velocity decreases to zero, the minimum lateral safety distance is calculated using the following formula. The lateral safety rule is generated based on the comparison between the minimum lateral safety distance and the lateral safety distance threshold. in, This refers to the lateral safety distance threshold. Let be the initial lateral velocity of the vehicle and the target vehicle in the Frenet coordinate system. The velocity value after moving with maximum lateral acceleration within the reaction time; This is the minimum lateral braking acceleration; When the interaction type is lane change interaction, the forward lane change safety rules and the backward lane change safety rules are calculated respectively according to the following formulas: in, For lane change duration, , The first one on the main lane The speed and longitudinal position of the vehicle with the obstacle. To ensure safe lane changing speed, This refers to the longitudinal position of the vehicle. If neither the forward lane change safety rule nor the backward lane change safety rule is met, it is determined that there is no safe and feasible lane change opportunity at present.

[0014] A second aspect of the present invention provides a risk assessment and security constraint generation apparatus, the apparatus comprising: The state acquisition module is configured to acquire perception information of the environment in which the autonomous vehicle is located, and convert the perception information into structured scene state information; The trajectory prediction module is configured to predict the future trajectories of each of the surrounding traffic participants in the prediction time domain based on the structured scene state information, combined with the road topology and lane connection relationship. During the prediction process, the prediction results are re-matched and corrected online according to the lateral offset and heading deviation of the surrounding traffic participants relative to the matched lanes, so as to obtain a predicted trajectory that conforms to the road constraints. The risk assessment module is configured to establish a potential collision damage model based on the motion state parameters of the vehicle and each of the surrounding traffic participants and the category information of the traffic participants within the prediction time domain, obtain the potential collision damage value of the vehicle and each of the surrounding traffic participants at each prediction time, and map the potential collision damage value to the real-time risk value at the current time through a spatiotemporal decay function. The rule generation module is configured to determine the maximum risk object based on the real-time risk value, determine the interaction type based on the spatial relationship between the vehicle and the maximum risk object, and generate safety rules including at least one of longitudinal safety rules, lateral safety rules and lane-changing safety rules based on the interaction type. The constraint generation module is configured to fuse the safety rules and output a set of safety constraints for the planning layer and the control layer. The output form of the safety constraint set includes at least one of the following: planning constraints for screening candidate trajectories, cost penalties for increasing the priority of safe trajectories in the cost function, and control triggers for triggering active safety actions.

[0015] A third aspect of the present invention provides an electronic device, comprising: at least one processor and at least one memory, wherein the memory stores program instructions or code; the program instructions or code are loaded and executed by the processor, causing the electronic device to implement the risk assessment and security constraint generation method as described in any of the preceding embodiments.

[0016] A fourth aspect of the present invention provides a storage medium having program instructions or code stored thereon, the program instructions or code being loaded and executed by a processor to implement the risk assessment and security constraint generation method as described in any of the preceding embodiments.

[0017] The beneficial effects of this invention include: By organically integrating road topology constraints, potential collision damage models, and spatiotemporal decay mechanisms, the risk assessment results are both forward-looking and physically interpretable. Compared with traditional methods that rely solely on collision time or fixed safety distances, this approach enables continuous characterization of potential conflicts in highly interactive and uncertain scenarios. By transforming risk assessment results into three output forms—planning constraints, cost penalties, and control triggers—in a hierarchical manner, safety constraints can not only affect candidate trajectory selection and cost function weighting at the planning layer, but also directly trigger active safety actions at the control layer when risks suddenly change. This significantly improves the traffic safety and decision-making rationality of autonomous vehicles in complex interactive scenarios on urban roads. The online rematching mechanism enables the predicted trajectory to respond promptly to the dynamic behaviors of traffic participants, such as lane changes, turns, and merging, avoiding the accumulation of prediction bias caused by lane deviation and improving the reliability of risk assessment input. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating a risk assessment and safety constraint generation method according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating the trajectory prediction and lane matching sub-process of another embodiment of the present invention; Figure 3 This is a flowchart illustrating the calculation of the comprehensive cost function and lane splicing in another embodiment of the present invention; Figure 4 This is a flowchart illustrating path correction and rematching based on Bézier curves, as shown in another embodiment of the present invention. Figure 5 The present invention provides the intersection prediction and simulation results of the vehicle and surrounding traffic participants as shown in the embodiments of the present invention. Figure 6 This is a flowchart illustrating a potential collision damage modeling process, as shown in another embodiment of the present invention. Figure 7 This is a flowchart illustrating the spatiotemporal decay function and real-time risk calculation sub-process of another embodiment of the present invention; Figure 8 The comparison results between this application and TTC are shown in the embodiments of this invention; Figure 9 This is a structural block diagram of a risk assessment and safety constraint generation device according to an embodiment of the present invention; Figure 10 This is a structural block diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0020] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0021] The following will clearly and completely describe the concept, specific structure, and technical effects of the present invention in conjunction with embodiments and accompanying drawings, so as to fully understand the purpose, features, and effects of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are all within the scope of protection of the present invention. Furthermore, all connections / linkages involved in the patent do not simply refer to direct contact between components, but rather to the ability to form a better connection structure by adding or reducing connecting accessories according to specific implementation conditions. The various technical features in this invention can be combined interactively without contradicting each other.

[0022] like Figure 1 As shown in the figure, an embodiment of the present invention provides a risk assessment and security constraint generation method, comprising: S1. Obtain perception information of the environment in which the autonomous vehicle is located, and convert the perception information into structured scene state information.

[0023] The perception information refers to the raw observation data about the vehicle's surrounding environment collected by sensors such as vehicle-mounted lidar, millimeter-wave radar, cameras, and combined positioning, including target location point clouds, target category image recognition results, target speed estimates, and the vehicle's own positioning and attitude information.

[0024] The structured scene state information includes: vehicle state information, surrounding traffic participant state information, traffic participant category information, and risk information for the current and historical moments. Specifically, the vehicle state information and traffic participant state information include position, speed, heading angle, longitudinal acceleration, and yaw rate; the traffic participant category information includes motor vehicle, non-motor vehicle, and pedestrian categories.

[0025] S2. Based on the structured scene state information, combined with the road topology and lane connection relationship, predict the future trajectory of each of the surrounding traffic participants in the prediction time domain. During the prediction process, online rematching correction is performed on the prediction results based on the lateral offset and heading deviation of the surrounding traffic participants relative to the matched lanes to obtain a predicted trajectory that conforms to the road constraints.

[0026] The road topology refers to the network of connections between lanes on the map through predecessor-successor relationships. Lane connectivity describes whether passage is permitted between any two adjacent lanes and their direction of travel. The prediction time domain is a fixed time window extending from the current moment into the future, used to depict the possible trajectories of traffic participants within this window. Online rematching correction refers to continuously checking whether the actual behavior of traffic participants still conforms to the previously matched lanes during the prediction process. Once the offset or heading deviation exceeds a set threshold, the process of reselecting matching lanes and regenerating the predicted trajectory is immediately triggered. Its purpose is to adapt to dynamic behaviors such as lane changes, entering intersections, or merging on ramps, ensuring that the predicted trajectory always remains consistent with the actual road structure.

[0027] S3. Within the prediction time domain, based on the motion state parameters of the vehicle and each of the surrounding traffic participants and the category information of the traffic participants, a potential collision damage model is established to obtain the potential collision damage values ​​of the vehicle and each of the surrounding traffic participants at each prediction time. The potential collision damage values ​​are then mapped to the real-time risk values ​​at the current time through a spatiotemporal decay function.

[0028] The potential collision damage model is a mathematical model that describes the degree of damage to persons or vehicles assumed to occur during a collision, based on parameters such as relative speed, velocity sum, and category sensitivity. The spatiotemporal decay function is a function that reduces the potential damage value based on the time difference between the predicted time and the current time, as well as the spatial distance between the traffic participant and their vehicle. This reflects the objective law that uncertainty increases with distance and the threat decreases with distance. In this scheme, the spatiotemporal decay function is constructed exponentially, causing the risk value to decrease smoothly with time and distance, thereby avoiding abrupt changes in the risk signal.

[0029] S4. Determine the maximum risk object based on the real-time risk value, and determine the interaction type based on the spatial relationship between the vehicle and the maximum risk object. Generate a safety rule based on the interaction type, including at least one of longitudinal safety rules, lateral safety rules, and lane-changing safety rules.

[0030] The highest-risk object refers to the object with the highest real-time risk value among all surrounding traffic participants. Prioritizing the generation of constraint rules for this object helps to focus on addressing the most severe potential conflicts when computational resources are limited. Spatial relationship refers to the lane relationship and relative longitudinal and lateral positional relationship between the vehicle and the highest-risk object. Interaction type is a semantic summary of this spatial relationship, specifically divided into longitudinal interaction within the same lane, lateral interaction between adjacent lanes, and lane-changing interaction in this scheme. Longitudinal safety rules are used to constrain following distance and braking distance, lateral safety rules are used to constrain lateral spacing and lateral movement trends, and lane-changing safety rules are used to constrain the longitudinal relationship between the vehicle in front and behind in the target lane and the vehicle itself. One or more of these three types of rules are called according to the interaction type.

[0031] S5. The safety rules are fused to output a set of safety constraints for the planning layer and the control layer. The output form of the safety constraint set includes at least one of the following: planning constraints for screening candidate trajectories, cost penalties for increasing the priority of safe trajectories in the cost function, and control triggers for triggering active safety actions.

[0032] Among these, planning constraints are applied directly as hard conditions to the candidate trajectory set, deleting candidate trajectories that do not meet safety rules. Cost penalties convert the degree of safety rule violation into an additional penalty term in the cost function, reducing the priority of unsafe trajectories without deleting them. Control triggers send direct action commands such as deceleration, braking, lane keeping, lane changing suppression, or emergency avoidance to the control layer when a safety rule is determined to be violated, bypassing the conventional planning process to achieve rapid response. The purpose of fusion is to eliminate potential conflicts between multiple rules and distribute them separately in three output forms: planning constraints, cost penalties, and control triggers.

[0033] Specifically, the method provided in this application continuously reads raw data from the perception system and performs spatiotemporal alignment. It then organizes the position, speed, heading angle, longitudinal acceleration, yaw rate, and category information of the vehicle and surrounding traffic participants into a unified data packet frame by frame for trajectory prediction. For each surrounding traffic participant, a predicted trajectory conforming to road constraints is generated based on its current motion state, combined with the road topology and lane connection relationships. During the prediction process, lateral offset and heading deviation are monitored in real time; if a threshold is exceeded, rematching and regenerating the predicted trajectory are triggered. Next, potential collision damage values ​​are calculated for each prediction time based on motion state parameters and category information, and the potential damage at each prediction time is converted into a real-time risk value at the current time using a spatiotemporal decay function. The maximum risk object is determined based on the real-time risk value, and longitudinal, lateral, or lane-changing safety rules are generated based on spatial relationships to determine the interaction type. Finally, the three types of rules are fused, and the safety constraint set is sent to the planning and control layers in three forms: planning constraints, cost penalties, and control triggers, achieving a closed loop from perception to execution.

[0034] The conversion of perceived information into structured scene state information can be achieved through a multi-sensor post-fusion approach based on timestamp alignment and coordinate system unification, merging LiDAR point cloud clustering results, camera visual detection results, and millimeter-wave radar tracking results into a unified target list. Alternatively, a pre-fusion approach can be used, extracting joint features from various sensor data at the feature level before outputting the target list. Online re-matching correction of predicted trajectories can be triggered using a lateral offset threshold-based method, where re-matching is triggered when the lateral distance between the actual target position and the centerline of the matching lane exceeds half the lane width. Alternatively, a heading deviation threshold-based method can be used, triggering re-matching when the difference between the target's heading and the matching lane's heading exceeds a set radian. The output form of the safety constraint set can employ a hard constraint method to directly delete candidate trajectories that violate safety rules, or a soft constraint method to reduce the priority of trajectories that violate safety rules by adding a penalty term to the cost function.

[0035] This embodiment organically integrates road topology constraints, potential collision damage models, and spatiotemporal decay mechanisms, enabling risk assessment results to possess both forward-looking and physically interpretable characteristics. Compared to traditional methods that rely solely on collision time or fixed safety distances, it can continuously characterize potential conflicts in highly interactive and uncertain scenarios. By hierarchically transforming risk assessment results into three output forms—planning constraints, cost penalties, and control triggers—safety constraints can not only influence candidate trajectory selection and cost function weighting at the planning layer but also directly trigger proactive safety actions at the control layer during sudden risk changes. This significantly improves the traffic safety and decision-making rationality of autonomous vehicles in complex urban road interaction scenarios. The online re-matching mechanism allows predicted trajectories to respond promptly to dynamic behaviors such as lane changes, turns, and merging by traffic participants, avoiding the accumulation of prediction biases caused by lane deviations and improving the reliability of risk assessment inputs.

[0036] In some embodiments, such as Figure 2 As shown, the prediction of future trajectories for each of the surrounding traffic participants in the prediction time domain includes: S201. For each of the surrounding traffic participants, based on their current motion state, a coarse predicted trajectory is generated by iteratively extrapolating using a constant rotation rate and constant acceleration model within the total predicted time, according to a preset departure step length, and the iterative extrapolation satisfies the following formula: in,( ) for the surrounding traffic participants in the first The predicted location coordinates at each time point , , , The surrounding traffic participants are respectively in the first The velocity, heading angle, longitudinal acceleration, and yaw rate at each predicted moment, where Δt is the distance from the step.

[0037] The constant revving and constant acceleration model is a kinematic model that assumes surrounding traffic participants maintain their current longitudinal acceleration and yaw rate within the prediction time domain. Its advantage is that it can recursively predict future position and attitude using only directly observable kinematic quantities such as current speed, heading angle, longitudinal acceleration, and yaw rate, without requiring additional driver intent modeling. The step length refers to the time increment used to evenly divide the prediction time domain into several equal time intervals, and the total prediction duration refers to the total span of the prediction time domain.

[0038] Coarse predicted trajectories refer to trajectory sequences obtained solely based on the aforementioned kinematic extrapolation, without yet being matched with the road structure.

[0039] Lane matching is performed based on the coarse predicted trajectory, and the matching methods include: S202. When the surrounding traffic participants first enter the prediction loop, global lane matching is performed on the pre-constructed map lane centerline point set to obtain the currently matched lane. S203. When the surrounding traffic participants have been continuously tracked, based on the lane matching result of the previous moment and the lane topology connection relationship, local lane matching is performed within the candidate range formed by the previous matched lane and its subsequent connected lanes to obtain the current matched lane.

[0040] Lane matching refers to the process of finding the most suitable assigned lane for surrounding traffic participants within the map's set of lane centerline points. Global lane matching involves searching for nearest neighbors or using a cost function on a pre-built set of lane centerline points across the entire map. This is suitable when surrounding traffic participants do not have historical matching information when they first enter the prediction cycle. Local lane matching searches only within the previously matched lane and its subsequent connected lanes determined by lane topology connections. Its purpose is to reduce the computational overhead of repeated global searches and improve matching stability by utilizing historical continuity. This is suitable when surrounding traffic participants are being continuously tracked. The currently matched lane refers to the assigned lane ultimately selected for the surrounding traffic participant within the current prediction cycle.

[0041] Specifically, for each surrounding traffic participant, the current position coordinates, velocity, heading angle, longitudinal acceleration, and yaw rate are first read. A coarse predicted trajectory sequence is then obtained through iterative extrapolation using a constant rotation rate and constant acceleration model based on a preset deviation step length Δt and the total prediction duration. Next, it is determined whether the surrounding traffic participant is entering the prediction loop for the first time. If it is, global lane matching is performed on a pre-constructed set of map lane centerline points to obtain the currently matched lane. If it has been continuously tracked, the lane matching result from the previous moment is read, and a candidate range consisting of the previous matched lane and its subsequent connected lanes is constructed based on lane topology connections. Local lane matching is then performed within this candidate range to obtain the currently matched lane. The coarse predicted trajectory and the current matched lane result serve as inputs for subsequent path correction and rematch triggering decisions.

[0042] This embodiment utilizes a constant rotation rate and constant acceleration model to recursively predict future short-term trajectories based solely on observable motion state variables at the current moment. This avoids the complex modeling process that relies on intent classification or large-scale training data, giving the prediction a significant advantage in real-time computation. By distinguishing between the initial entry into the prediction loop and the continuous tracking scenario, global lane matching and local lane matching are employed respectively. This allows for global lane assignment determination when a new target appears, while utilizing historical results to reduce redundant searches for continuously tracked targets. Overall, it balances matching accuracy and computational efficiency, providing stable lane assignment input for subsequent trajectory correction based on road constraints.

[0043] In some embodiments, such as Figure 3 As shown, the process of obtaining the predicted trajectory that conforms to the road constraints based on the currently matched lane includes: S2011. Calculate the distance cost and heading cost between the end point of the coarse predicted trajectory and the candidate lane, respectively, and construct a comprehensive cost function: in, Candidate lanes The overall cost, , These are the distance weighting coefficient and the heading weighting coefficient, respectively. The Euclidean distance between two points. The endpoint of the coarsely predicted trajectory. Candidate lanes The center point, For the coarse trajectory end heading, To match lane heading; S2012. Select the candidate lane that corresponds to the minimum value of the comprehensive cost as the target lane; S2013. When the length of a single target lane is insufficient to support the complete predicted distance, the target lane and its subsequent connecting lanes are centerline-stitched according to the lane topology connection relationship to form a continuous reference path. : In this context, ⊕ indicates that the centerline is spliced ​​according to the lane connection relationship. The number of subsequent lane layers to meet the predicted length requirements.

[0044] Specifically, in obtaining the coarse predicted trajectory endpoint... and its terminal heading Then, several candidate lanes near the area where the target traffic participant is located are retrieved from the map, and the distance cost from the coarse prediction endpoint to the center point of the candidate lane is calculated for each candidate lane using the Euclidean distance formula. The heading cost from the coarsely predicted final heading to the candidate lane heading is calculated using the formula based on the absolute value of the angle difference. The comprehensive cost value of each candidate lane is calculated according to the comprehensive cost function. The comprehensive cost values ​​of each candidate lane are compared, and the candidate lane with the minimum comprehensive cost is selected as the target lane. Then, the remaining length of the target lane along the lane direction and the predicted trajectory length L are calculated. k The relationship is that when the remaining length of the target lane is less than the predicted trajectory length L k At that time, the target lane and its subsequent connecting lanes are sequentially connected according to the lane topology. , Until By splicing the centerlines end to end, a continuous reference path R with a length that meets the prediction requirements is obtained. k .

[0045] Among them, the distance weighting coefficient w d With heading weighting coefficient w θ The values ​​can be determined by segmenting the calibration according to road type, increasing the distance weight in straight road scenarios and increasing the heading weight in curved road scenarios. Alternatively, an online adaptive adjustment based on historical prediction errors can be used, with the actual deviation at the prediction end serving as feedback to dynamically correct the weight ratio. The centerline stitching of continuous reference paths can be achieved by directly connecting the discrete points of the map's lane centerlines end-to-end, or by using spline curves for a smooth transition near the stitching point to avoid abrupt changes in heading at the stitching point.

[0046] This embodiment incorporates both distance and heading costs into the candidate lane selection process by introducing a comprehensive cost function. This avoids the bias caused by relying solely on the closest distance or the best heading, ensuring that the target lane selection result possesses both spatial proximity and directional consistency. By introducing continuous reference paths and stitching centerlines according to lane topology connections, even in complex road structures with limited single-lane lengths, such as intersections, forks, and ramp merging, the predicted trajectory can still obtain sufficiently long reference path support. This avoids trajectory correction failures due to insufficient reference paths and improves prediction stability under complex road structures.

[0047] In some embodiments, such as Figure 4 As shown, obtaining the predicted trajectory conforming to road constraints based on the continuous reference path includes: S2014. Taking the current position of the surrounding traffic participants as the starting point and the nearest matching point of the predicted trajectory length on the continuous reference path as the ending point, a cubic Bézier curve is used to correct the predicted trajectory. The cubic Bézier curve is a parameterized curve defined by four control points P0, P1, P2, and P3 to obtain the predicted trajectory that conforms to the road constraints. The cubic Bézier curve satisfies: in, The current location of the surrounding traffic participants. , Control points of the Bézier curve To predict trajectory length In the continuous reference path the nearest matching point on, , These are the tangent vectors at the starting and ending points, respectively; The path correction process ensures that the generated predicted trajectory is consistent with the target's current direction of motion at the starting point, consistent with the direction of the continuous reference path at the ending point, and smoothly transitions in the middle section, thus preserving both the target's short-term motion trend and conforming to the road geometry.

[0048] S2015. During the prediction process, the lateral offset and heading consistency of the surrounding traffic participants relative to the currently matched lane are monitored in real time. When one of the following conditions is met, the re-matching of the currently matched lane and the regeneration of the predicted trajectory are triggered: in, The width of the currently matched lane. The heading of the currently matched lane. Provides real-time location and heading information for surrounding traffic participants.

[0049] Wherein, the lateral offset refers to the target's current position p k The minimum Euclidean distance to the center line of the currently matched lane, w v* The width of the currently matched lane is used; 0.5 times the lane width can be used as the boundary for determining whether the target has deviated to the lane edge. Heading deviation refers to the target's current heading θ. k Current matching lane heading θ vk* The absolute value of the difference is processed by the wrap function, which normalizes the angle difference to the interval (-π, π). When any of the above conditions are met, it is considered that the target has deviated from the original matching lane, and a new matching lane needs to be selected and the predicted trajectory regenerated based on the new matching lane.

[0050] Specifically, in obtaining the current location P0 of the target traffic participant and the continuous reference path R k Then, set P3 as the predicted trajectory length L. k In R kThe nearest matching point is used to determine the endpoint P3, and two intermediate control points are also determined. Then, the parameter λ is uniformly discretized in the interval (0,1), and the corrected coordinates of the predicted trajectory are calculated point-by-point using the cubic Bézier curve formula, resulting in a predicted trajectory sequence that conforms to road constraints. During the prediction process, the current target position p is continuously read. k With heading θ k Calculate the lateral offset and match it with 0.5w vk* The heading deviation is calculated and compared with the heading deviation threshold of 0.5. Once either condition is met, the original matched lane is immediately cleared, lane matching is re-executed, and the predicted trajectory is regenerated based on the new matched lane.

[0051] The calibration of control points for cubic Bézier curves can be achieved by using a scaling factor of 0.3 times the predicted trajectory length as the tangent vector scaling factor, ensuring a moderate influence range of the tangent direction near the start and end points. Alternatively, the scaling factor can be adaptively adjusted based on lane curvature, appropriately reducing the factor on lanes with high curvature to prevent excessive control point deviation. The rematching trigger condition can employ a single threshold triggering method, using only whether the lateral offset exceeds 0.5 times the lane width as the sole criterion. Alternatively, a parallel determination method using lateral offset and heading deviation can be used, triggering rematching upon either condition being met to enhance responsiveness to lane changes and steering maneuvers.

[0052] This embodiment uses a cubic Bézier curve to correct the coarse predicted trajectory, ensuring that the predicted trajectory retains the short-term movement trend of the target at the starting point and conforms to the direction of the continuous reference path at the ending point. The curve is smooth and second-order continuous, avoiding the abrupt changes in heading caused by directly splicing polylines. By introducing a dual rematching trigger condition based on lateral offset and heading deviation, the prediction process can respond promptly to dynamic behaviors such as lane changes, entry into intersections, or ramp merging, avoiding the prediction errors accumulated by extrapolating along the original matched lane over a long period. This ensures that the prediction results maintain high road consistency in dynamic scenarios. Figure 5 The figure shows the intersection prediction and simulation results between the vehicle and the surrounding traffic participants.

[0053] In some embodiments, such as Figure 6 As shown, the establishment of the potential collision damage model to obtain the potential collision damage values ​​between the vehicle and each of the surrounding traffic participants at each predicted time includes: S301, within the prediction time domain, for the first... The surrounding traffic participants, in the first The motion state parameters of the vehicle and the surrounding traffic participants are obtained at each predicted time point. These motion state parameters include the vehicle's motion state parameters at the predicted time point. The speed at each predicted time point The surrounding traffic participants in the first The speed at each predicted time point And the two in the Relative velocity at each predicted time point .

[0054] The higher the relative velocity value, the greater the difference in momentum between the two objects at the moment of collision, and the stronger the collision impact. (Vehicle speed v) e (k) relative to the speeds v of surrounding traffic participants n (k) together are used to characterize the overall operating energy level of the system in a collision scenario. The larger the sum of the two, the greater the overall kinetic energy carried by the collision system and the more severe the collision consequences.

[0055] S302. Based on the motion state parameters, construct the comprehensive velocity characterization term according to the following formula. : in, This is the speed difference weighting coefficient, with a value range of (0,1).

[0056] Among them, the comprehensive speed characterization term The relative velocity term and the velocity sum term are weighted and fused using a velocity difference weighting coefficient γ. When it is necessary to emphasize the impact intensity of the collision, the value of γ is increased to amplify the influence of the relative velocity term; when it is necessary to emphasize the total energy level of the collision system, the value of γ is decreased to amplify the influence of the velocity sum term. This comprehensive velocity characterization avoids the bias caused by evaluating collision consequences solely based on relative velocity or solely based on velocity sum as a single indicator, enabling damage quantification to have good discriminative power under different collision modes.

[0057] S303. Perform a nonlinear mapping on the comprehensive velocity characterization term to obtain the basic damage quantity. : in, , These are nonlinear mapping parameters.

[0058] In this system, α controls the amplification factor at the input of the logarithmic function, and β ensures that the input of the logarithmic function is always positive to avoid divergence, and can also be used to distinguish different driving styles. The logarithmic mapping method maintains the basic damage amount... With the comprehensive velocity characterization term While increasing monotonically, it suppresses the phenomenon of damage values ​​amplifying too quickly in high-speed scenarios, so that the damage model has a relatively stable numerical distribution in different speed ranges, which facilitates the unified calibration of subsequent risk thresholds.

[0059] Specifically, damage assessments are performed sequentially for each surrounding traffic participant within the prediction time domain. For the nth surrounding traffic participant at the k-th prediction time, the vehicle's speed at the k-th prediction time is first read from the trajectory prediction results. and the speed of the nth surrounding traffic participant at the kth predicted time. The relative velocities of the two are calculated using the vector difference method. The relative velocity term is then weighted and fused with the velocity sum term to obtain the comprehensive velocity characterization term. Further analysis of the comprehensive velocity characterization term The basic damage amount is obtained by performing a logarithmic mapping. This baseline damage amount serves as the input parameter for subsequently introducing class weights and spatiotemporal decay.

[0060] Among them, relative velocity The calculation can be performed using a vector difference method, considering the difference between the two-dimensional velocity vectors of the vehicle and surrounding traffic participants, and taking their magnitudes to accurately reflect the intensity of relative motion under lateral and longitudinal coupling conflicts. Alternatively, a projection method along the connecting line can be used, considering only the difference in the velocity components of both parties along the connecting line, suitable for pure longitudinal following scenarios. The calibration of the nonlinear mapping parameters α and β can be performed offline by fitting the damage distribution of typical collision cases, using the least squares method to determine a set of parameter values ​​suitable for a general driving style. Alternatively, a tiered calibration method can be used, selecting different combinations of parameter values ​​for conservative, standard, and aggressive driving styles.

[0061] This embodiment integrates relative speed and speed, two complementary indicators, by introducing a comprehensive speed characterization term. This allows damage quantification results to be no longer limited to a single speed indicator, and enables the output of physically reasonable damage values ​​under different conflict scenarios such as rear-end collisions, lateral cut-ins, and lateral interactions. By performing a logarithmic mapping on the comprehensive speed characterization term to obtain the basic damage quantity, damage values ​​are prevented from amplifying too rapidly in high-speed scenarios. The numerical distribution of the damage model is more stable, facilitating subsequent decision-making using a unified risk threshold under different road conditions. Overall, this improves the stability and interpretability of the risk assessment results.

[0062] In some embodiments, such as Figure 7 As shown, mapping the potential collision damage value to the real-time risk value at the current moment using a spatiotemporal decay function includes: S304, Based on the traffic participant category information, the first... The weighting coefficients of the matching categories of the surrounding traffic participants Construct the first formula as follows The surrounding traffic participants in the first Potential collision damage value at each predicted time : The category weight coefficient ωn refers to the pre-assigned weight values ​​based on the traffic participant category (motor vehicle, non-motor vehicle, or pedestrian), used to reflect the differences in injury risk faced by different types of objects under the same collision conditions. Under the same comprehensive speed characteristic... In this case, the category weight coefficients for vulnerable road users such as pedestrians and non-motorized vehicles are significantly greater than those for motor vehicles, resulting in higher potential collision damage values. A higher value indicates a higher priority in risk assessment.

[0063] S305. Constructing time decay and spatial decay functions using exponential functions: in, , Control the rate of decay of time risk and the rate of decay of spatial risk respectively. Let these be the spatial decay function and the temporal decay function. For the current moment, For the first The surrounding traffic participants in the first Distance from the predicted time position to the predicted trajectory endpoint; This represents the difference between the current time and the predicted terminal time.

[0064] λ in the time decay function t Used to control the rate of time risk decay, kt is the time difference between the predicted time and the current time. The larger the time difference, the farther away the predicted time is from the current time. Since the prediction error increases with time, the corresponding risk contribution should be smaller.

[0065] λ in the spatial decay function d Used to control the rate of decay of space risk For the first The surrounding traffic participants in the first The distance from the predicted time point to the predicted trajectory endpoint is considered. The greater the distance, the smaller the threat to the current vehicle, and the smaller the corresponding risk contribution should be. Using an exponential function to construct the decay function ensures a smooth and continuous decay process, avoiding abrupt changes in the risk signal.

[0066] S306. The real-time risk value at the current moment is calculated based on the potential collision damage value, the time decay function, and the spatial decay function.

[0067] Real-time risk value refers to the potential collision damage value at each predicted time. According to the corresponding time decay weight w T (kt) and spatial decay weight w D ( The scalar value is obtained by weighting and summing the values. This value reflects the overall risk level posed by the nth surrounding traffic participant to the vehicle at the current moment, after considering the increase in time uncertainty and the decay of spatial distance.

[0068] Specifically, the baseline damage amount of the nth surrounding traffic participant at the kth predicted time is obtained. Then, based on the traffic participant category information, the corresponding category weight coefficient ωn is read from a pre-established category weight lookup table, and the potential collision damage value is calculated. Then, the time decay weight of the k-th prediction time relative to the current time is calculated according to the time decay function, and the spatial decay weights of surrounding traffic participants and the vehicle at the k-th prediction time are calculated according to the spatial decay function. This applies to all prediction times within the prediction time domain. Press the corresponding w T with w D A weighted sum is performed to obtain the real-time risk value of the nth surrounding traffic participant at the current moment. This real-time risk value serves as the input for subsequent determination of the highest risk object and interaction type.

[0069] The category weight coefficient ωn can be determined by classifying traffic participants according to their categories, assigning incremental values ​​to motor vehicles, non-motor vehicles, and pedestrians to reflect the higher injury sensitivity of vulnerable traffic participants. Alternatively, a dynamic lookup table method based on finer-grained type classification can be used, assigning different weight values ​​to subcategories such as cyclists, electric two-wheelers, and child pedestrians. The decay rates λt and λd of the time decay function and spatial decay function can be determined offline based on simulation results of typical scenarios, obtaining unified values ​​by fitting the risk value curves as a function of time and distance under different scenarios. Alternatively, an online adjustment method can be used based on the predicted time domain length and environmental characteristics, increasing the decay rate in low-speed urban scenarios and decreasing it in high-speed scenarios.

[0070] This embodiment explicitly incorporates the differences in damage sensitivity among different types of traffic participants into the risk quantification process by introducing category weighting coefficients. This makes the risk assessment results more sensitive to vulnerable traffic participants such as pedestrians and non-motorized vehicles, avoiding the problem of underestimating the risk of vulnerable groups caused by treating all traffic participants equally. By introducing time decay and spatial decay functions based on exponential functions, the potential damage value at each predicted time can be smoothly discounted to the current time according to its time difference and spatial distance from the current time. Overall, the risk value has both forward-looking and immediate availability, and the exponential form ensures that the risk signal changes continuously and smoothly, facilitating stable safety decisions made by the planning and control layers based on the risk value.

[0071] In some embodiments, the interaction type includes at least one of longitudinal interaction in the same lane, lateral interaction in adjacent lanes, and lane-changing interaction; When the interaction type is longitudinal interaction within the same lane, the minimum longitudinal safety distance is calculated based on the worst-case assumption using the following formula. The longitudinal safety rule is generated based on the minimum longitudinal safety distance: in, For reaction time, For maximum acceleration, For minimum braking deceleration, For maximum braking deceleration, , These are the speeds of the following vehicle and the vehicle in front, respectively. , These are the lengths of the rear vehicle and the front vehicle, respectively.

[0072] The worst-case scenario assumption refers to assuming that the vehicle in front immediately decelerates at its maximum braking speed at the current moment. Braking is applied while the following vehicle continues to accelerate at maximum speed within the reaction time ρ. After the reaction time has elapsed, the following vehicle will decelerate with minimal braking. Extreme operating conditions requiring braking until complete stop. Minimum longitudinal safety distance. This refers to the initial longitudinal distance required to ensure that the following vehicle can stop before the preceding vehicle stops at its position under this operating condition.

[0073] When the interaction type is lateral interaction between adjacent lanes, in the Frenet coordinate system or the lane local coordinate system, based on the worst-case lateral approach process where the vehicle and the maximum risk object move towards each other with maximum lateral acceleration during the reaction time, and both decelerate with minimum lateral braking acceleration after the reaction time ends until their lateral velocity decreases to zero, the minimum lateral safety distance is calculated using the following formula. The lateral safety rule is generated based on the comparison between the minimum lateral safety distance and the lateral safety distance threshold. in, This refers to the lateral safety distance threshold. Let be the initial lateral velocity of the vehicle and the target vehicle in the Frenet coordinate system. The velocity value after moving with maximum lateral acceleration within the reaction time; This is the minimum lateral braking acceleration.

[0074] The Frenet coordinate system refers to a curved coordinate system established with the lane centerline as the reference axis. Its advantage lies in decoupling the longitudinal displacement along the lane direction from the lateral displacement perpendicular to the lane direction, making the calculation of lateral safety rules more intuitive.

[0075] When the interaction type is lane change interaction, the forward lane change safety rules and the backward lane change safety rules are calculated respectively according to the following formulas: in, For lane change duration, , The first one on the main lane The speed and longitudinal position of the vehicle with the obstacle. For safety, change lanes and speed.

[0076] Among them, the forward lane change safety rule refers to the judgment criterion that when a vehicle intends to change lanes in front of a vehicle in the target lane, assuming that the vehicle behind in the target lane is traveling at the maximum permissible acceleration, the vehicle performs the lane change according to the predetermined lane change trajectory and the expected speed, and the vehicle behind in the target lane and the vehicle always meet the longitudinal safety conditions throughout the entire lane change process.

[0077] The rear lane change safety rule refers to the requirement that when a vehicle intends to change lanes behind a vehicle in the target lane, the longitudinal distance between the vehicle and the vehicle in front in the target lane after the lane change still meets the minimum longitudinal safety distance. The criteria for judgment.

[0078] If neither the forward lane change safety rule nor the backward lane change safety rule is met, it is determined that there is no safe and feasible lane change opportunity at present.

[0079] The statement that there is no safe and feasible lane-changing opportunity means that during the entire process from the current decision-making moment to the completion of the lane change, it is impossible to complete the lane change in front of or behind a vehicle in the target lane. In this case, the safety rules module outputs a lane-changing suppression command to the planning layer to prevent the vehicle from forcibly changing lanes with insufficient lane-changing gaps.

[0080] Specifically, the interaction type is determined based on the object of greatest risk and its spatial relationship with the vehicle. When the interaction type is determined to be longitudinal interaction within the same lane, the minimum longitudinal safety distance is calculated based on the worst-case scenario. and the actual longitudinal distance between the vehicle and the vehicle in front and By comparing the results, longitudinal safety rules are derived. When the interaction type is determined to be a lateral interaction between adjacent lanes, the minimum lateral safety distance is calculated in the Frenet coordinate system. The lateral safety distance threshold is then compared to obtain the lateral safety rules. When the interaction type is determined to be a lane-changing interaction, the forward lane-changing safety rules and the backward lane-changing safety rules are calculated and judged respectively. If neither of them is satisfied, the judgment result that there is no safe and feasible lane-changing opportunity is output.

[0081] Among them, the minimum longitudinal safety distance The parameter calibration can be preset in stages according to vehicle type and road type, with different settings for passenger cars and commercial vehicles in urban and highway scenarios. , , The value of ρ can be adjusted online based on road weather and road surface adhesion coefficient to improve the conservatism of the minimum braking deceleration estimate under low adhesion conditions. For the lateral safety rules, the coordinate system can be selected using the Frenet coordinate system to decouple the longitudinal and lateral movements along the lane direction and perpendicular to the lane direction, respectively. Alternatively, a lane local coordinate system can be used to simplify the calculation by employing an approximate Cartesian coordinate system on straight sections or sections with large curvature radii.

[0082] This embodiment models three types of safety rules—longitudinal, lateral, and lane-changing—based on interaction type and derives parameters according to worst-case assumptions. This ensures that the generated safety rules maintain a safety margin even under the most stringent opponent behavior within the reaction time, exhibiting significantly higher robustness compared to methods that estimate collision time solely based on the current observed speed. By introducing a two-way verification mechanism for forward and backward lane-changing safety rules and actively abandoning lane changes when neither is satisfied, vehicles avoid forced lane changes with insufficient lane-changing gaps, improving overall traffic safety in dense traffic scenarios. Furthermore, it provides the planning and control layers with executable constraints with clear physical meaning, such as… Figure 8 The figure shows a comparison between the method (risk) of this application embodiment and TTC.

[0083] like Figure 9 As shown in the figure, a risk assessment and safety constraint generation device 60 provided in this application embodiment includes: The state acquisition module 601 is configured to acquire perception information of the environment in which the autonomous vehicle is located, and convert the perception information into structured scene state information; The trajectory prediction module 602 is configured to predict the future trajectories of each of the surrounding traffic participants in the prediction time domain based on the structured scene state information, combined with the road topology and lane connection relationship. During the prediction process, the prediction results are re-matched and corrected online according to the lateral offset and heading deviation of the surrounding traffic participants relative to the matched lanes, so as to obtain a predicted trajectory that conforms to the road constraints. The risk assessment module 603 is configured to establish a potential collision damage model based on the motion state parameters of the vehicle and each of the surrounding traffic participants and the category information of the traffic participants within the prediction time domain, obtain the potential collision damage value of the vehicle and each of the surrounding traffic participants at each prediction time, and map the potential collision damage value to the real-time risk value at the current time through a spatiotemporal decay function. The rule generation module 604 is configured to determine the maximum risk object based on the real-time risk value, determine the interaction type based on the spatial relationship between the vehicle and the maximum risk object, and generate safety rules including at least one of longitudinal safety rules, lateral safety rules and lane-changing safety rules based on the interaction type. The constraint generation module 605 is configured to fuse the safety rules and output a set of safety constraints for the planning layer and the control layer. The output form of the safety constraint set includes at least one of the following: planning constraints for screening candidate trajectories, cost penalties for increasing the priority of safe trajectories in the cost function, and control triggers for triggering active safety actions.

[0084] It should be noted that the risk assessment and safety constraint generation device provided in the above embodiments is only illustrated by the division of the above functional modules when performing risk assessment and safety constraints. In actual applications, the above functions can be assigned to different functional modules as needed. That is, the internal structure of the risk assessment and safety constraint generation device will be divided into different functional modules to complete all or part of the functions described above.

[0085] Furthermore, the risk assessment and safety constraint generation apparatus and the risk assessment and safety constraint generation method provided in the above embodiments belong to the same concept, and the specific way in which each module performs its operation has been described in detail in the method embodiments, and will not be repeated here.

[0086] Please see Figure 10 This invention provides an electronic device 4000, which includes at least one processor 4001 and at least one memory 4003.

[0087] The data interaction between the processor 4001 and the memory 4003 can be achieved through at least one communication bus 4002. This communication bus 4002 may include a path for transmitting data between the processor 4001 and the memory 4003. The communication bus 4002 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. The communication bus 4002 can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 10 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0088] Optionally, the electronic device 4000 may further include a transceiver 4004, which can be used for data interaction between the electronic device and other electronic devices, such as sending and / or receiving data. It should be noted that in practical applications, the transceiver 4004 is not limited to one type, and the structure of the electronic device 4000 does not constitute a limitation on the embodiments of the present invention.

[0089] Processor 4001 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this invention. Processor 4001 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0090] The memory 4003 may be a ROM (Read Only Memory) or other type of static storage device capable of storing static information and instructions, RAM (Random Access Memory) or other type of dynamic storage device capable of storing information and instructions, or an EEPROM (Electrically Erasable Programmable Read Only Memory), CD-ROM (Compact Disc Read Only Memory) or other optical disc storage (including CDs, laser discs, digital versatile optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program instructions or code in the form of instructions or data structures and accessible by the electronic device 4000, but not limited thereto.

[0091] The memory 4003 stores program instructions or code, and the processor 4001 can read the program instructions or code stored in the memory 4003 through the communication bus 4002.

[0092] When the program instructions or code are executed by the processor 4001, the risk assessment and security constraint generation methods in the above embodiments are implemented.

[0093] Furthermore, this embodiment of the invention provides a storage medium storing program instructions or code, which is loaded and executed by a processor to implement the risk assessment and security constraint generation method described above.

[0094] This invention provides a computer program product, which includes program instructions or code stored in a storage medium. The processor of an electronic device reads the program instructions or code from the storage medium, loads and executes the program instructions or code, enabling the electronic device to implement the risk assessment and security constraint generation method described above.

[0095] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of the present invention.

Claims

1. A method for risk assessment and security constraint generation, characterized in that, include: S1. Obtain perception information of the environment in which the autonomous vehicle is located, and convert the perception information into structured scene state information; The structured scene state information includes the vehicle's state, the state of surrounding traffic participants, traffic participant category information, and risk information for the current and historical moments. S2. Based on the structured scene state information, combined with the road topology and lane connection relationship, predict the future trajectory of each of the surrounding traffic participants in the prediction time domain, and in the prediction process, perform online rematching correction on the prediction results according to the lateral offset and heading deviation of the surrounding traffic participants relative to the matched lanes, so as to obtain the predicted trajectory that meets the road constraints. S3. Within the prediction time domain, based on the motion state parameters of the vehicle and each of the surrounding traffic participants and the category information of the traffic participants, a potential collision damage model is established to obtain the potential collision damage value of the vehicle and each of the surrounding traffic participants at each prediction time, and the potential collision damage value is mapped to the real-time risk value at the current time through a spatiotemporal decay function. S4. Determine the maximum risk object based on the real-time risk value, and determine the interaction type based on the spatial relationship between the vehicle and the maximum risk object. Generate a safety rule including at least one of longitudinal safety rules, lateral safety rules and lane-changing safety rules based on the interaction type. S5. The safety rules are fused to output a set of safety constraints for the planning layer and the control layer. The output form of the safety constraint set includes at least one of the following: planning constraints for screening candidate trajectories, cost penalties for increasing the priority of safe trajectories in the cost function, and control triggers for triggering active safety actions.

2. The risk assessment and security constraint generation method according to claim 1, characterized in that, The prediction of future trajectories for each of the surrounding traffic participants in the prediction time domain includes: S201. For each of the surrounding traffic participants, based on their current motion state, a coarse predicted trajectory is generated by iteratively extrapolating using a constant rotation rate and constant acceleration model within the total predicted time, according to a preset departure step length, and the iterative extrapolation satisfies the following formula: in,( ) for the surrounding traffic participants in the first The predicted location coordinates at each time point , , , The surrounding traffic participants are respectively in the first The velocity, heading angle, longitudinal acceleration, and yaw rate at each predicted moment, where Δt is the distance from the walk; Lane matching is performed based on the coarse predicted trajectory, and the matching methods include: S202. When the surrounding traffic participants first enter the prediction loop, global lane matching is performed on the pre-constructed map lane centerline point set to obtain the currently matched lane. S203. When the surrounding traffic participants have been continuously tracked, based on the lane matching result of the previous moment and the lane topology connection relationship, local lane matching is performed within the candidate range formed by the previous matched lane and its subsequent connected lanes to obtain the current matched lane.

3. The risk assessment and security constraint generation method according to claim 2, characterized in that, The process of obtaining the predicted trajectory that conforms to the road constraints based on the currently matched lane includes: S2011. Calculate the distance cost and heading cost between the end point of the coarse predicted trajectory and the candidate lane, respectively, and construct a comprehensive cost function: in, Candidate lanes The overall cost, , These are the distance weighting coefficient and the heading weighting coefficient, respectively. The Euclidean distance between two points. The endpoint of the coarsely predicted trajectory. Candidate lanes The center point, For the coarse trajectory end heading, To match lane heading; S2012. Select the candidate lane that corresponds to the minimum value of the comprehensive cost as the target lane; S2013. When the length of a single target lane is insufficient to support the complete predicted distance, the target lane and its subsequent connecting lanes are centerline-stitched according to the lane topology connection relationship to form a continuous reference path. : In this context, ⊕ indicates that the centerline is spliced ​​according to the lane connection relationship. To meet the predicted length requirements for the number of subsequent lane layers, To match lane line information in the lane.

4. The risk assessment and security constraint generation method according to claim 3, characterized in that, The predicted trajectory conforming to road constraints is obtained based on the continuous reference path, including: S2014. Taking the current position of the surrounding traffic participants as the starting point and the nearest matching point of the predicted trajectory length on the continuous reference path as the ending point, a cubic Bézier curve is used to correct the predicted trajectory to obtain the predicted trajectory that conforms to the road constraints. The cubic Bézier curve satisfies: in, The current location of the surrounding traffic participants. , Control points of the Bézier curve To predict trajectory length In the continuous reference path the nearest matching point on, , These are the tangent vectors at the starting and ending points, respectively; S2015. During the prediction process, the lateral offset and heading consistency of the surrounding traffic participants relative to the currently matched lane are monitored in real time. When one of the following conditions is met, the re-matching of the currently matched lane and the regeneration of the predicted trajectory are triggered: in, The width of the currently matched lane. The heading of the currently matched lane. Provides real-time location and heading information for surrounding traffic participants.

5. The risk assessment and security constraint generation method according to claim 1, characterized in that, The establishment of a potential collision damage model yields the potential collision damage values ​​between the vehicle and each of the surrounding traffic participants at each predicted time, including: S301, within the prediction time domain, for the first... The surrounding traffic participants, in the first The motion state parameters of the vehicle and the surrounding traffic participants are obtained at each predicted time point. These motion state parameters include the vehicle's motion state parameters at the predicted time point. The speed at each predicted time point , No. The surrounding traffic participants in the first The speed at each predicted time point And the two in the Relative velocity at each predicted time point ; S302. Based on the motion state parameters, construct the comprehensive velocity characterization term according to the following formula. : in, The speed difference weighting coefficient; S303. Perform a nonlinear mapping on the comprehensive velocity characterization term to obtain the basic damage quantity. : in, , These are nonlinear mapping parameters.

6. The risk assessment and security constraint generation method according to claim 5, characterized in that, The process of mapping the potential collision damage value to the real-time risk value at the current moment using a spatiotemporal decay function includes: S304, Based on the traffic participant category information, the first... The weighting coefficients of the matching categories of the surrounding traffic participants Construct the first formula as follows The surrounding traffic participants in the first Potential collision damage value at each predicted time : S305. Constructing time decay and spatial decay functions using exponential functions: in, , Control the rate of decay of time risk and the rate of decay of spatial risk respectively. Let these be the spatial decay function and the temporal decay function. For the current moment, For the first The surrounding traffic participants in the first Distance from the predicted time position to the predicted trajectory endpoint; The difference between the current time and the predicted terminal time. S306. The real-time risk value at the current moment is calculated based on the potential collision damage value, the time decay function, and the spatial decay function.

7. The risk assessment and safety constraint generation method according to claim 1, characterized in that, The interaction types include at least one of longitudinal interaction within the same lane, lateral interaction between adjacent lanes, and lane-changing interaction. When the interaction type is longitudinal interaction within the same lane, the minimum longitudinal safety distance is calculated based on the worst-case assumption using the following formula. The longitudinal safety rule is generated based on the minimum longitudinal safety distance: in, For reaction time, For maximum acceleration, For minimum braking deceleration, For maximum braking deceleration, , These are the speeds of the following vehicle and the vehicle in front, respectively. , These are the lengths of the rear vehicle and the front vehicle, respectively. When the interaction type is lateral interaction between adjacent lanes, in the Frenet coordinate system or the lane local coordinate system, based on the worst-case lateral approach process where the vehicle and the maximum risk object move towards each other with maximum lateral acceleration during the reaction time, and both decelerate with minimum lateral braking acceleration after the reaction time ends until their lateral velocity decreases to zero, the minimum lateral safety distance is calculated using the following formula. The lateral safety rule is generated based on the comparison between the minimum lateral safety distance and the lateral safety distance threshold. in, This refers to the lateral safety distance threshold. Let be the initial lateral velocity of the vehicle and the target vehicle in the Frenet coordinate system. The velocity value after moving with maximum lateral acceleration within the reaction time; This is the minimum lateral braking acceleration; When the interaction type is lane change interaction, the forward lane change safety rules and the backward lane change safety rules are calculated respectively according to the following formulas: in, For lane change duration, , The first one on the main lane The speed and longitudinal position of the vehicle with the obstacle. To ensure safe lane changing speed, This refers to the longitudinal position of the vehicle. If neither the forward lane change safety rule nor the backward lane change safety rule is met, it is determined that there is no safe and feasible lane change opportunity at present.

8. A risk assessment and safety constraint generation device, characterized in that, The device includes: The state acquisition module is configured to acquire perception information of the environment in which the autonomous vehicle is located, and convert the perception information into structured scene state information; The trajectory prediction module is configured to predict the future trajectories of each of the surrounding traffic participants in the prediction time domain based on the structured scene state information, combined with the road topology and lane connection relationship. During the prediction process, the prediction results are re-matched and corrected online according to the lateral offset and heading deviation of the surrounding traffic participants relative to the matched lanes, so as to obtain a predicted trajectory that conforms to the road constraints. The risk assessment module is configured to establish a potential collision damage model based on the motion state parameters of the vehicle and each of the surrounding traffic participants and the category information of the traffic participants within the prediction time domain, obtain the potential collision damage value of the vehicle and each of the surrounding traffic participants at each prediction time, and map the potential collision damage value to the real-time risk value at the current time through a spatiotemporal decay function. The rule generation module is configured to determine the maximum risk object based on the real-time risk value, determine the interaction type based on the spatial relationship between the vehicle and the maximum risk object, and generate safety rules including at least one of longitudinal safety rules, lateral safety rules and lane-changing safety rules based on the interaction type. The constraint generation module is configured to fuse the safety rules and output a set of safety constraints for the planning layer and the control layer. The output form of the safety constraint set includes at least one of the following: planning constraints for screening candidate trajectories, cost penalties for increasing the priority of safe trajectories in the cost function, and control triggers for triggering active safety actions.

9. An electronic device, characterized in that, include: At least one processor and at least one memory, wherein program instructions or code are stored in the memory; the program instructions or code are loaded and executed by the processor, causing the electronic device to implement the risk assessment and security constraint generation method as described in any one of claims 1 to 7.

10. A storage medium storing program instructions or code thereon, characterized in that, The program instructions or code are loaded and executed by the processor to implement the risk assessment and security constraint generation method as described in any one of claims 1 to 7.

Images