CN-121973800-A - Automatic driving planning method and system for roundabout scene uncertainty modeling

CN121973800ACN 121973800 ACN121973800 ACN 121973800ACN-121973800-A

Abstract

The invention discloses an automatic driving planning method and system for roundabout scene uncertainty modeling, wherein the method comprises the steps of obtaining roundabout scene information and constructing vectorized representation of the roundabout scene information; the method comprises the steps of inputting vectorized representation of the roundabout scene information into a detection head to obtain uncertainty parameters of the roundabout scene information, establishing an initial self-vehicle query vector, updating the initial self-vehicle query vector by combining the uncertainty parameters and vectorized representation of the roundabout scene information to obtain an updated self-vehicle query vector, generating an initial predicted running path according to the updated self-vehicle query vector, and optimizing the initial predicted running path through uncertainty constraint to generate an automatic driving running path. According to the method, the uncertainty parameters are built for the roundabout scene, so that the self-vehicle query vector is generated, the prediction driving path is generated, and the robustness and the stability of track prediction in the roundabout scene are improved aiming at the problem of the distribution randomness of vehicles and static obstacles in the complex roundabout scene.

Inventors

YIN ZHISHUAI
XIE YING
Zhai Xukai
TANG CHENGCHEN
ZHAO FENGYUN
JIANG LONGFEI
QIN YEFEI

Assignees

武汉理工大学

Dates

Publication Date: 20260505
Application Date: 20260127

Claims (10)

1. An automatic driving planning method for modeling of a roundabout scene uncertainty is characterized by comprising the following steps: Acquiring rotary island scene information and constructing vectorized representation of the rotary island scene information, wherein the rotary island scene information comprises rotary island vehicles and rotary island barriers; Calculating a Laplace distribution probability density function according to vectorization representation of the roundabout scene information to obtain roundabout scene uncertainty parameters considering uncertainty; establishing an initial self-vehicle query vector, and updating the initial self-vehicle query vector by combining the roundabout scene uncertainty parameter and the vectorized representation of the roundabout scene information to obtain an updated self-vehicle query vector; Generating an initial predicted running path according to the updated self-vehicle query vector, optimizing the initial predicted running path through uncertainty constraint with environmental understanding to generate a final automatic driving running path, wherein the uncertainty constraint with environmental understanding specifically calculates the distance between the self-vehicle and the roundabout scene according to the roundabout scene uncertainty parameter, and adjusts the predicted running path according to the calculation result.
2. The automatic driving planning method for modeling of roundabout scene uncertainty according to claim 1, wherein the constructing of the vectorized representation of the roundabout scene information is specifically: Acquiring multi-view images of a vehicle in running, extracting BEV image features from the images, and fusing the BEV image features of different time frames to obtain a BEV feature map; Respectively learning the characteristics of the rotary island vehicle and the characteristics of the rotary island obstacle from the BEV characteristic diagram by utilizing a deformable attention mechanism to obtain a learned rotary island vehicle query vector and a rotary island obstacle query vector; And decoding the rotary island vehicle query vector and the rotary island obstacle query vector by using the multi-layer perceptron to obtain the vectorized representation of the rotary island vehicle and the rotary island obstacle.
3. The method for automatically driving planning for modeling the roundabout scene uncertainty according to claim 1, wherein obtaining the uncertainty parameters of the roundabout scene considering the uncertainty comprises obtaining the uncertainty parameters of the roundabout vehicle and the uncertainty parameters of the roundabout obstacle; the position uncertainty parameter of the rotary island vehicle is specifically obtained by the following steps: Decomposing BEV projection center coordinates of the vectorization representation of the rotary island vehicle into an x component and a y component, respectively adopting Laplace distribution to carry out uncertainty modeling to obtain a Laplace distribution probability density function of the rotary island vehicle position, and taking a Laplace distribution mean value and a scale parameter of the rotary island vehicle position as position uncertainty parameters; The track uncertainty parameter of the rotary island vehicle is specifically obtained by the following modes: And carrying out uncertainty modeling on the future motion driving path track of the rotary island vehicle by using the Laplace distribution to obtain a Laplace distribution probability density function of the rotary island vehicle track, and taking the Laplace distribution mean value and the scale parameter of the Laplace distribution probability density function as track uncertainty parameters.
4. The automatic driving planning method for modeling of the roundabout scene uncertainty according to claim 3, wherein the uncertainty parameter of the roundabout obstacle is specifically an observed position uncertainty parameter of the roundabout obstacle, and is specifically obtained by the following method: and obtaining the vertex coordinates of the observation position of the roundabout obstacle, carrying out uncertainty modeling by adopting Laplace distribution, obtaining a Laplace distribution probability density function of the observation point of the roundabout obstacle, and taking the Laplace distribution mean value and the scale parameter of the Laplace distribution probability density function as the uncertainty parameter of the observation position.
5. The method for automatically driving planning for modeling the roundabout scene uncertainty according to claim 1, wherein updating the initial vehicle query vector specifically comprises updating a vehicle query vector for interaction between a vehicle and the roundabout vehicle and a vehicle query vector for interaction between the vehicle and the roundabout obstacle; the vehicle query vector of interaction between the vehicle and the rotary island vehicle is specifically obtained by the following steps: The vectorization representation of the rotary island vehicles is used as keys and values of a transducer, and the initial self-vehicle query vectors, the keys and the values are calculated through a deterministic attention mechanism to obtain fixed weights of the self-vehicle and the rotary island vehicles; Carrying out change distribution calculation on the fixed weight and the position uncertainty parameter of the rotary island vehicle to obtain the random weight of the interaction between the self-vehicle and the rotary island vehicle, and multiplying the random weight by the vectorized representation of the rotary island vehicle to obtain the self-vehicle query vector of the interaction between the self-vehicle and the rotary island vehicle; the vehicle query vector of interaction between the vehicle and the roundabout obstacle is specifically obtained by the following steps: The vectorization representation of the roundabout barrier is used as a key and a value of a transducer, and a self-vehicle query vector, a key and a value of the self-vehicle interaction with the roundabout barrier are calculated through a deterministic attention mechanism, so that a fixed weight of the self-vehicle interaction with the roundabout barrier is obtained; And carrying out change distribution calculation on the fixed weight and the uncertainty parameters of the observed positions of the rotary island obstacles to obtain the random weight of the vehicle obstacles of the vehicle and the rotary island, multiplying the random weight by the vectorized representation of the rotary island obstacles to obtain the vehicle query vector of the interaction of the vehicle and the rotary island obstacles, and taking the vehicle query vector as the updated vehicle query vector.
6. The method for automatically planning the driving by modeling the roundabout scene uncertainty according to claim 1, wherein the initial predicted driving path of the own vehicle is obtained by decoding the updated inquiry vector of the own vehicle through a multi-layer perceptron decoder, and the uncertainty constraint with environmental understanding comprises an uncertainty collision constraint of the own vehicle and the roundabout vehicle and an uncertainty boundary constraint of the own vehicle and the roundabout obstacle.
7. The automatic driving planning method for the roundabout scene uncertainty modeling according to claim 6, wherein the uncertainty collision constraint of the host vehicle and the roundabout vehicle is specifically that the relative distance between the host vehicle and the roundabout vehicle is calculated according to the track uncertainty parameter of the roundabout vehicle, the probability that the relative distance between the host vehicle and the roundabout vehicle is larger than a first safety distance threshold is calculated as a first safety probability, the magnitude of the first safety probability and the magnitude of the first confidence threshold are judged, and when the first safety probability is smaller than the first confidence threshold, the difference between the first confidence threshold and the first safety probability is used as a collision constraint loss to adjust the multi-layer sensor parameter, so that the optimized predicted driving path is obtained.
8. The automatic driving planning method for the roundabout scene uncertainty modeling according to claim 6, wherein the uncertainty boundary constraint of the host vehicle and the roundabout obstacle is specifically that Euclidean distance between the host vehicle and the boundary of the roundabout obstacle is calculated according to the observed position uncertainty parameter of the roundabout obstacle, the probability that the Euclidean distance between the host vehicle and the boundary of the roundabout obstacle is larger than a second safety distance threshold is calculated as second safety probability, the magnitude of the second safety probability and the second confidence threshold is judged, and when the second safety probability is smaller than the second confidence threshold, the difference value between the second confidence threshold and the second safety probability is used as boundary constraint loss to adjust the multi-layer sensor parameter, so that an optimized predicted driving path is obtained.
9. An automated driving planning system for modeling of roundabout scene uncertainty, the system comprising: the system comprises a vectorization representation module, a vector representation module and a storage module, wherein the vectorization representation module is used for acquiring rotary island scene information and constructing vectorization representation of the rotary island scene information, and the rotary island scene information comprises rotary island vehicles and rotary island barriers; the uncertainty parameter acquisition module is used for calculating a Laplace distribution probability density function according to the vectorized representation of the roundabout scene information to obtain roundabout scene uncertainty parameters considering uncertainty; The self-vehicle query vector updating module is used for establishing an initial self-vehicle query vector, and updating the initial self-vehicle query vector by combining the vector representation of the roundabout scene uncertainty parameter and the roundabout scene information to obtain an updated self-vehicle query vector; The system comprises a driving path prediction and optimization module, a prediction and optimization module and a prediction and optimization module, wherein the driving path prediction and optimization module is used for generating an initial prediction driving path according to the updated self-vehicle query vector, optimizing the initial prediction driving path through uncertainty constraint with environmental understanding, and generating a final automatic driving path, wherein the uncertainty constraint with environmental understanding is used for calculating the distance between a self-vehicle and a roundabout scene according to uncertainty parameters of the roundabout scene information, and adjusting the prediction driving path according to a calculation result.
10. A computer storage medium, characterized in that a computer program executable by a processor is stored therein, which computer program performs the roundabout scenario-oriented uncertainty modeling autopilot planning method according to any one of claims 1-8.

Description

Automatic driving planning method and system for roundabout scene uncertainty modeling Technical Field The invention relates to the technical field of automatic driving path planning, in particular to an automatic driving planning method and system for modeling of roundabout scene uncertainty. Background The existing automatic driving movement planning method is mainly based on a regularized road structure, and relies on clear geometric constraint and explicit traffic semantic reasoning driving paths. However, in complex urban scenarios, such as a roundabout scenario in urban roads, the host needs to complete real-time interactions with the roundabout entrance, the roundabout vehicles within the roundabout, and the roundabout exit (e.g., vehicles with varying speeds and travel paths within the roundabout), and the roundabout obstacles (e.g., temporarily placed traffic cones) within a limited space and time. In the scene, speed difference and interval fluctuation generally exist between the self-vehicle and the rotary island vehicle, the traffic flow state is changed frequently, meanwhile, the position distribution of the rotary island vehicle and the rotary island obstacle has stronger randomness, and the uncertainty of the environmental understanding result is further increased. When the self-vehicle runs in the urban rotary island scene, the multi-source uncertainty caused by environmental change and multi-vehicle interaction is needed to be simultaneously dealt with, and the planning difficulty of the vehicle running path is obviously higher than that of the conventional structured road scene. Aiming at the uncertainty problem faced in the running process of the vehicle, part of traditional methods are explored in the conventional structured roads. For example, one class of methods models uncertainty in an upstream environmental understanding stage, constrains travel path generation by introducing risk assessment or constructing a risk field when the environment is understood, thereby improving the safety of planning. However, such methods focus only on the uncertainty measure at the environmental understanding level, failing to cover the uncertain propagation of the environmental understanding of the planning whole process. Another class of methods introduces uncertainty modeling in the downstream travel path generation stage, such as generating multiple candidate travel paths to mitigate the risk of environmental understanding errors, but such methods do not take into account the problem of uncertainty coupling between the environmental understanding stage and the planning stage as well. Therefore, when facing complex urban scenes, such as a roundabout scene in an urban road, the conventional motion planning method still lacks a planning framework capable of systematically modeling multi-source uncertainty and combining safety and robustness. Disclosure of Invention In order to solve the problem of insufficient research on the roundabout scene uncertainty in the prior art, the invention provides an automatic driving planning method and system for modeling the roundabout scene uncertainty, so as to realize the generation of a predicted driving path with both safety and robustness. The technical scheme adopted by the invention is as follows: The method for automatically planning driving for the roundabout scene uncertainty modeling comprises the following steps: Acquiring rotary island scene information and constructing vectorized representation of the rotary island scene information, wherein the rotary island scene information comprises rotary island vehicles and rotary island barriers; Calculating a Laplace distribution probability density function according to vectorization representation of the roundabout scene information to obtain roundabout scene uncertainty parameters considering uncertainty; establishing an initial self-vehicle query vector, and updating the initial self-vehicle query vector by combining the roundabout scene uncertainty parameter and the vectorized representation of the roundabout scene information to obtain an updated self-vehicle query vector; Generating an initial predicted running path according to the updated self-vehicle query vector, optimizing the initial predicted running path through uncertainty constraint with environmental understanding to generate a final automatic driving running path, wherein the uncertainty constraint with environmental understanding specifically calculates the distance between the self-vehicle and the roundabout scene according to the roundabout scene uncertainty parameter, and adjusts the predicted running path according to the calculation result. According to the scheme, the vectorization characterization of the information of the rotary island scene is constructed specifically as follows: Acquiring multi-view images of a vehicle in running, extracting BEV image features from the images, and fusing the BEV image features of different time fram