CN-121328675-B - Tree-like retrieval enhanced automatic driving automobile compliance behavior induction method

CN121328675BCN 121328675 BCN121328675 BCN 121328675BCN-121328675-B

Abstract

The invention provides a tree-like retrieval enhanced automatic driving automobile compliance behavior induction method, which belongs to the technical field of automatic driving and comprises the steps of 1) reading a traffic regulation text, partitioning the traffic regulation text according to a paragraph structure, further subdividing a partitioning result by using a large language model to obtain a plurality of traffic regulation clauses 2) constructing a road traffic scene tree structure and a traffic regulation clause scene semantic analysis model to form a traffic regulation knowledge base 3) extracting scene elements based on driving data, acquiring node distribution of each element in the tree structure, constructing query characterization, matching corresponding coding clauses 4) in the traffic regulation knowledge base to construct a behavior constraint thinking chain, generating a forbidden behavior list and a mandatory behavior list, and guiding an automatic driving automobile to run in compliance. The invention constructs a road traffic scene tree structure, provides a tree search algorithm, can synchronously search for the universality rule and the pertinence rule, and improves the search and behavior constraint generation precision of the traffic rule.

Inventors

YU RONGJIE
Jian Bowen

Assignees

同济大学

Dates

Publication Date: 20260512
Application Date: 20250926

Claims (10)

1. A tree-like search enhanced automated driving vehicle compliance generalization method, comprising the steps of: S1, reading a traffic regulation text, blocking the text according to a paragraph structure, further subdividing a blocking result by using a large language model to obtain a plurality of traffic regulation clauses, and ensuring that a single clause corresponds to a single road traffic scene; S2, constructing a road traffic scene tree structure by referring to a road-infrastructure-traffic management-traffic participant-environment-information scene classification six-layer model, representing various scene elements and hierarchical relations thereof, generating codes corresponding to the elements by applying a full-path coding method, constructing a traffic rule scene semantic analysis model by utilizing a large language model prompting word fine tuning technology, automatically labeling scene elements related to each rule, acquiring corresponding element codes, and forming a traffic rule knowledge base in a traffic rule and element coding form; S3, extracting static scene elements by combining vehicle position information and a high-precision map, identifying dynamic scene elements from vehicle-end sensing data by utilizing a detection algorithm, acquiring node distribution of each element in a tree structure, constructing query characterization by utilizing element codes of a current node and ancestor nodes thereof, and matching corresponding coding clauses in a traffic rule knowledge base; and S4, constructing a behavior constraint thinking chain, judging the behavior constraint types in each search term, extracting a behavior description text, and inducing and generating a forbidden behavior list and a mandatory behavior list according to the behavior constraint types to guide the autonomous vehicle to run in compliance.
2. The tree search enhanced automatic driving car compliance generalization method of claim 1, wherein in S1, the paragraph structure comprises a line feed and a rule term specific identifier.
3. The method for inducing the compliance of the automatic driving automobile with the enhanced tree search according to claim 1, wherein in the step S1, the method is further subdivided specifically in that through prompt word engineering, a large language model judges whether a given text segment contains a plurality of road traffic scenes, if so, the text segment is further subdivided, otherwise, the text segment is not processed, and finally, each term is ensured to correspond to only a single traffic scene.
4. The tree search enhanced automated driving automobile compliance generalization method of claim 1, wherein the scene classification six-layer model comprises ISO 34504, ASAM OpenDRIVE, ASAM OpenSCENARIO documents.
5. The tree search enhanced automatic driving automobile compliance generalization method of claim 1, wherein S2 specifically comprises the steps of: S21, representing the hierarchical relationship among different elements by utilizing a multi-way tree, wherein the multi-way tree nodes are in one-to-one correspondence with scene elements, the deeper the node depth in the multi-way tree is, the lower the hierarchy of the corresponding scene elements is, and the finer the granularity of the scene information is contained; s22, utilizing a full path coding method to endow independent codes for all scene elements in the S21 tree; S23, initializing a trainable soft prompt word for each scene element, taking the soft prompt word, the scene element name and the rule term text as input signals of a large language model, taking whether the rule term comprises the element as an output signal of a label, and learning the soft prompt word corresponding to each scene element end to end by utilizing the fine adjustment of the prompt word to train an objective function, wherein the objective function is expressed as: ; ; ; Wherein, the Is the first A personal rule clause text; is the first Names of individual scene elements; for its corresponding soft-cue word, wherein, For the predefined soft stop word length, set to 20, Embedding layer dimensions for a large language model; for pre-training weights of Is a large language model of (a); is the first The provision of the individual compliance is at the first True values at individual scene elements; Predicting a probability distribution for a next token for the large language model, wherein, Representing the size of the vocabulary; Is a cross entropy loss function; s24, after the objective function training in S23 is completed, giving traffic regulation clauses, executing model reasoning node by node, obtaining all scene elements associated with the traffic regulation clauses, and reading corresponding codes; s25, repeating S24 until all traffic regulation clauses are marked, and finally forming a structured traffic regulation knowledge base of the traffic regulation clause-element codes.
6. The method for summarizing the compliance behavior of an automatic driving automobile with enhanced tree search according to claim 1, wherein in S2, the full-path encoding method is specifically a tree search path from a root node to the end of a current node.
7. The tree search enhanced automatic driving car compliance generalization method of claim 1, wherein said S3 comprises the steps of: S31, reading road information of a place where a vehicle is located from a map, and extracting static scene element characteristics including road type, facility type, lane exclusive right, lane width, plane linearity, longitudinal section characteristics, one-way lane number, front lane number change, road passing direction, surrounding built environment, traffic sign, traffic marking and road central isolation facility setting conditions; s32, based on a vehicle forward video, identifying weather states, visibility levels and surrounding traffic participants in real time by using a detection algorithm to obtain dynamic scene elements; S33, acquiring node distribution of each element in a tree structure, extracting element codes of each node and ancestor nodes thereof, and combining to obtain query characterization; and S34, traversing each term in the traffic regulation knowledge base in the S25, and if the scene element combination corresponding to the term belongs to the query characterization, recording the scene element combination into a retrieval result to finally form a complete traffic regulation list.
8. The method for enhancing the compliance of an automated driving vehicle with tree search of claim 7, wherein said detecting algorithm comprises identifying surrounding traffic participants using Yolo algorithm, identifying weather conditions using convolutional neural network and estimating visibility level in S32.
9. The tree search enhanced automatic driving automobile compliance generalization method of claim 1, wherein S4 specifically comprises the steps of: S41, given a text of traffic regulation clause, constructing a large language model thinking chain reasoning prompt word, judging the behavior constraint type contained in the clause, and extracting vehicle behavior description on the basis; s42, repeating the step S41 to process and retrieve one by one to obtain traffic regulations; s43, according to the behavior constraint types, summarizing description texts of the forbidden behaviors and the mandatory behaviors respectively, removing repeated or synonymous behavior descriptions, and summarizing and generating a forbidden behavior list and a mandatory behavior list.
10. The tree search enhanced automated driving vehicle compliance generalization method of claim 9, wherein in S41, the behavior constraint type comprises mandatory behavior, forbidden behavior, and authorized behavior.

Description

Tree-like retrieval enhanced automatic driving automobile compliance behavior induction method Technical Field The invention belongs to the technical field of automatic driving, and particularly relates to a tree-shaped retrieval enhanced automatic driving automobile compliance induction method. Background Compliance driving is an important precondition for automated driving of automobiles into traffic systems and for large-scale applications. The traditional automatic driving automobile adopts a rule algorithm, and by summarizing human driving experience, decision logic is designed, and manual writing rules are supported to adapt to traffic regulation constraints. However, the existing automatic driving system adopts a data driving algorithm, and utilizes an end-to-end model with strong fitting capability to directly learn the mapping relation from sensor data to control signals, so that the internal mechanism of the model is difficult to analyze, and the compliance is difficult to ensure. Therefore, there is a need for accurate retrieval of scene-related traffic regulations, generalizing vehicle behavior constraints, and providing support for data-driven algorithm training and optimization. The existing retrieval technology is mainly based on the similarity relation between query characterization and knowledge documents, measures the similarity by utilizing keywords "Li B,Wang Y,Mao J,et al.Driving everywhere with large language model policy adaptation[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.2024:14948-14957" or natural language semantic features "Lewis P,Perez E,Piktus A,et al.Retrieval-augmented generation for knowledge-intensive nlp tasks[J].Advances in neural information processing systems,2020,33:9459-9474", and selects the knowledge documents with higher similarity from a knowledge base as a retrieval result. However, driving scene types are various and dynamic change, the hierarchical relationship among scene elements is complex, and only the similarity relationship between the current scene and the legal scene is measured, so that the universality of legal regulations is easy to miss, and further the restriction of the vehicle behavior is not fully induced. Disclosure of Invention The invention aims to solve the problems of low traffic regulation retrieval precision and incomplete behavior constraint induction, and provides a tree retrieval enhanced automatic driving automobile compliance behavior induction method, which is characterized by comprising the following steps: S1, reading a traffic regulation text, blocking the text according to a paragraph structure, further subdividing a blocking result by using a large language model to obtain a plurality of traffic regulation clauses, and ensuring that a single clause corresponds to a single road traffic scene; S2, constructing a road traffic scene tree structure by referring to a road-infrastructure-traffic management-traffic participant-environment-information scene classification six-layer model, representing various scene elements and hierarchical relations thereof, generating codes corresponding to the elements by applying a full-path coding method, constructing a traffic rule scene semantic analysis model by utilizing a large language model prompting word fine tuning technology, automatically labeling scene elements related to each rule, acquiring corresponding element codes, and forming a traffic rule knowledge base in a traffic rule and element coding form; S3, extracting static scene elements by combining vehicle position information and a high-precision map, identifying dynamic scene elements from vehicle-end sensing data by utilizing a detection algorithm, acquiring node distribution of each element in a tree structure, constructing query characterization by utilizing element codes of a current node and ancestor nodes thereof, and matching corresponding coding clauses in a traffic rule knowledge base; s4, constructing a behavior constraint thinking chain, judging the behavior constraint types in each search term, extracting a behavior description text, and inducing and generating a forbidden behavior list and a mandatory behavior list according to the behavior constraint types to guide the automatic driving vehicle to run in compliance. Further, in S1, the paragraph structure includes a linefeed and a regulatory term specific identifier (e.g., "first bar"). Further, in S1, the method is characterized in that through prompt word engineering, a large language model judges whether a given text segment contains a plurality of road traffic scenes, if so, the text segment is further subdivided, otherwise, the text segment is not processed, and finally, each term is ensured to correspond to only a single traffic scene. Further, the scene classification six-layer model comprises ISO 34504, ASAM OpenDRIVE, ASAM OpenSCENARIO files. Further, S2 specifically includes the following steps: S21, representin