CN-121989932-A - Traffic risk decision method and system integrating traffic rule constraint and Zhou Che tracks

CN121989932ACN 121989932 ACN121989932 ACN 121989932ACN-121989932-A

Abstract

The invention discloses a driving risk decision method and a driving risk decision system for fusing traffic rule constraint and Zhou Che tracks, wherein the method comprises the steps of predicting future tracks of surrounding vehicles, calculating dynamic risk field force through a constructed driving risk vector field model based on the predicted tracks, quantifying traffic rules into equivalent constraint field force, configuring constraint switch variables capable of being independently regulated and controlled to control the effective state of the equivalent constraint field force, carrying out dynamic risk partitioning according to risk field force distribution and real-time motion state, carrying out risk avoidance decision based on quantitative comparison results of the risk field force and the constraint field force at all moments in the future, triggering an emergency risk avoidance mode if the risk force exceeds the constraint force, selectively breaking through rules through adjusting the switch variables, and generating the risk avoidance tracks, and otherwise keeping in the rule compliance mode. The invention realizes the spanning from passive response to active prediction and from regular rigid compliance to intelligent flexible trade-off, and remarkably improves the safety decision capability of the automatic driving automobile under a complex scene.

Inventors

ZHENG XUNJIA
YAO PI
JIANG JUNHAO
CHEN XING
LIU HUI

Assignees

重庆文理学院

Dates

Publication Date: 20260508
Application Date: 20260203

Claims (10)

1. A traffic risk decision method integrating traffic rule constraint and Zhou Che tracks is characterized by comprising the following steps: s1, acquiring historical track information of surrounding vehicles, and predicting the motion track of each surrounding vehicle in a preset future time window based on the acquired historical track information; S2, calculating and obtaining dynamic risk field force used for representing collision risk of surrounding vehicles to the own vehicle at each prediction moment in a future time window through a pre-constructed running risk vector field model based on the motion trail obtained in the step S1 and the real-time relative motion state of the own vehicle and the surrounding vehicles; S3, quantifying at least one traffic rule into an equivalent constraint field force for representing the constraint effect on the vehicle, and configuring constraint switch variables capable of being regulated and controlled independently for each type of traffic rule so as to control whether the equivalent constraint field force of the corresponding traffic rule is effective or not by switching the states of the constraint switch variables; s4, dynamically defining a plurality of areas with different risk grades based on the real-time relative motion states of the own vehicle and surrounding vehicles by taking the own vehicle as a center based on the spatial distribution of the dynamic risk field force calculated in the step S2, and forming dynamic risk zones; And S5, carrying out risk avoidance decision based on the dynamic risk field force at each prediction time in the future time window provided by the step S2, the equivalent constraint field force provided by the step S3 and the dynamic risk partition provided by the step S4, wherein the method comprises the steps of triggering an emergency risk avoidance mode if the dynamic risk field force at the prediction time in the future time window exceeds the equivalent constraint field force, and otherwise, keeping in a rule compliance mode.
2. The method according to claim 1, wherein in step S1, a neural network prediction model incorporating a spatiotemporal attention mechanism is provided for trajectory prediction, comprising: encoding the history track information of surrounding vehicles by means of an encoder to obtain a corresponding history hidden state sequence; Processing the history hidden state sequence, and calculating the interaction influence weight of each related surrounding vehicle on the own vehicle based on the relative distance, the relative speed and the vehicle size information between the vehicles through a spatial attention mechanism to obtain a spatial attention weighted feature vector; Screening historical time sequence characteristics with strong influence on current track prediction through a time attention mechanism, and calculating corresponding importance weights of the historical time sequence characteristics to obtain time attention weighted characteristic vectors; and generating the motion trail of each surrounding vehicle in a preset future time window through decoding the spatial attention weighted feature vector and the time attention weighted feature vector.
3. The method of claim 2, wherein the neural network prediction model is a long-short-term memory network STA-LSTM model based on a spatiotemporal feature attention mechanism.
4. The method according to claim 1, wherein the dynamic risk field force is a vector and the direction is set to be the relative speed direction of the own vehicle and the corresponding surrounding vehicle, and the driving risk vector field model is set to be based on a piecewise defined functional relation reflecting the nonlinear change of the risk with the distance, and the distance on which the driving risk vector field model is based is set to be an effective relative distance combined with a dynamic gradient adjustment coefficient, wherein the dynamic gradient adjustment coefficient is configured to be capable of adaptively adjusting the weight of the contribution of the longitudinal and transverse distances to the risk according to the relative motion state of the own vehicle and the surrounding vehicle.
5. The method of claim 4, wherein the dynamic gradient adjustment factor is configured to automatically decrease the longitudinal dynamic gradient adjustment factor when the host vehicle is traveling in the same direction as the surrounding vehicles and there is a rear-end collision risk situation, such that the calculated value of the effective relative distance is decreased, and such that the calculated dynamic risk field force is increased.
6. The method according to claim 1, wherein in step S3, the constraint switch variables have an effective state and a failure state, and the constraint switch variables are in the effective state initially so that the corresponding traffic rules generate equivalent constraint field forces with constraint effects, and when the dynamic risk field forces at the future prediction time exceeds the equivalent constraint field forces, at least part of the constraint switch variables corresponding to the traffic rules are switched to the failure state so that the equivalent constraint field forces of the corresponding traffic rules are zeroed.
7. The method according to claim 6, characterized in that in step S5 the triggering condition of the emergency avoidance mode is set such that, within a preset future time window, there is at least one prediction instant whose scalar value of the dynamic risk field force exceeds the scalar value of the equivalent constraint field force scaled by the safety factor; The emergency risk avoiding mode is set to comprise the following steps: Identifying the type of the risk area to which the risk threat belongs according to the risk partition result determined in the step S4; Selecting a traffic rule to be broken through according to a preset mapping strategy which is used for associating different risk area types with a breakthroughable rule base; a control instruction is sent, and in step S3, the constraint switch variable corresponding to the selected traffic rule to be broken through is switched to a failure state; And generating a corresponding risk avoiding track.
8. The method according to claim 1, wherein in step S4, the dynamic risk zone comprises at least a longitudinal high risk zone, a side scratch risk zone and a safety zone, wherein longitudinal boundary values of the longitudinal high risk zone are dynamically calculated based on a speed of the vehicle, a heading angle, a displacement within a preset reaction time and a relative braking distance between surrounding vehicles having a risk of collision with the vehicle; the relative braking distance is set to be calculated depending on the real-time relative speed of the own vehicle and the surrounding vehicles and the estimated relative deceleration.
9. The method of claim 1, wherein the quantified traffic rules include at least traffic lights, road boundaries, and crosswalk lines.
10. A system for implementing the traffic risk decision method of fusing traffic rule constraints with Zhou Che trajectories according to any one of claims 1-9, comprising: a motion trajectory prediction unit configured to acquire history trajectory information of surrounding vehicles and predict motion trajectories of the respective surrounding vehicles within a preset future time window based on the acquired history trajectory information; The risk field acquisition unit is configured to calculate and acquire dynamic risk field force used for representing collision risk of surrounding vehicles to the own vehicle at each prediction moment in a future time window through a pre-constructed driving risk vector field model based on the acquired motion trail and the real-time relative motion state of the own vehicle and the surrounding vehicles; The system comprises a constraint field acquisition unit, a constraint field generation unit and a constraint field generation unit, wherein the constraint field acquisition unit is configured to quantize at least one traffic rule into an equivalent constraint field force for representing the constraint action on a vehicle, and configure constraint switch variables capable of being independently regulated and controlled for each type of traffic rule so as to control whether the equivalent constraint field force of the corresponding traffic rule is effective or not by switching the states of the constraint switch variables; The risk partition unit is configured to dynamically define a plurality of areas with different risk levels based on the real-time relative motion states of the own vehicle and surrounding vehicles by taking the own vehicle as a center based on the spatial distribution of the acquired dynamic risk field force, so as to form a dynamic risk partition; The risk avoidance decision unit is configured to perform risk avoidance decision based on the dynamic risk field force, the equivalent constraint field force and the dynamic risk partition at each prediction time in the future time window, and comprises triggering an emergency risk avoidance mode if the dynamic risk field force at the prediction time in the future time window exceeds the equivalent constraint field force, and otherwise maintaining in a rule compliance mode.

Description

Traffic risk decision method and system integrating traffic rule constraint and Zhou Che tracks Technical Field The invention relates to the field of environment perception, decision and planning of automatic driving automobiles, in particular to a traffic risk decision method and system integrating traffic rule constraint and Zhou Che tracks, which can integrate high-precision track prediction and traffic rule quantification constraint in a complex dynamic traffic environment to carry out prospective and refined traffic risk assessment and intelligent risk avoidance decision. Background With the rapid development of automatic driving technology and the gradual promotion of commercial landing, ensuring the driving safety of vehicles in various complex, variable and even extreme traffic scenes has become the most core and challenging subject in the field. The driving risk is accurately and real-time identified, and the driving behavior which is safe, reasonable and accords with the regulations is generated on the basis, so that the driving risk is a key for preventing traffic accidents and improving the reliability and social acceptance of automatic driving. Currently, a mainstream risk assessment and decision method in industry generally has significant technical bottlenecks and limitations when dealing with a multiple constraint environment in which dynamic traffic participants (surrounding vehicles, also simply referred to as "surrounding vehicles") and static traffic rules are fused. The risk assessment methods widely adopted at present can be mainly classified into the following categories, and the following categories have the defects: The first category is evaluation methods based on classical kinematic indexes, such as collision time (Time to Collision, TTC), headway (TIME HEADWAY, THW), etc. Such methods are computationally simple, but are essentially based on a "snapshot" assessment of the current or instantaneous historical state. The core problem is that the look-ahead is seriously insufficient. The risk is based on 'collision possibly happening in the future', while methods such as TTC and the like assume that the vehicle keeps a current motion state, and cannot effectively predict and cope with future risk changes of the vehicle caused by actions such as acceleration, deceleration, steering and the like. This results in the system often triggering a response only when the risk is imminent, early warning is delayed, and failure may occur due to insufficient response in high speed scenarios. The second type is a Potential Field theory (Potential Field) based method. The method is to construct a repulsive field and a gravitational field for the obstacle and the target point respectively, and the vehicle moves under the action of resultant force. However, the conventional artificial potential field method is mostly a static scalar field, and only can indicate that a certain point is at risk, but cannot characterize the directionality of the risk. Meanwhile, the field force parameters are set empirically, so that dynamic interaction influence between vehicles due to different relative speeds and different masses is difficult to accurately describe. More importantly, the existing potential field method generally models traffic rules (such as lane lines and signal lamps) simply as insurmountable 'rigid virtual barriers', and lacks a mechanism for uniformly quantifying and intelligently balancing dynamic risks. The third class is a data-driven trajectory prediction and probabilistic risk assessment method. Such methods (e.g., using deep learning models such as LSTM, GNN, etc.) are typically in a split state with subsequent risk assessment, decision modules, although they can predict future trajectories of surrounding vehicles, even output uncertainties. The prediction module takes the trajectory position error minimization as an optimization target, and the risk module carries out independent evaluation based on the prediction result. This "predictive-assessment" tandem architecture may result in the system over focusing on the "most likely" to occur trajectories, while ignoring the "low probability, high risk" trajectories, with potential safety hazards. The fourth class is a rule and logic based decision architecture, finite State Machine (FSM), behavioral Tree (BT). The method encodes the traffic rules into clear logic conditions, and has clear structure. But its fundamental disadvantage is the regular "stiffness". In the face of extremely high risk scenes such as 'out of control rear-end collision' of a rear vehicle, the system falls into ethical and technical double dilemmas that the strict adherence to the rule of 'stop of a red light' leads to collision, and the system lacks a quantitative decision basis to judge whether, when and how to temporarily and conditionally break through a certain rule constraint in exchange for higher safety (for example, accelerating through the red light crossing to