An adaptive autonomous driving decision-making method fusing user preferences and traffic rules

Abstract

The application relates to a kind of adaptive automatic driving decision-making methods of fusing user preference and traffic rules, comprising the following steps: extracting automatic driving scene information and user instruction;Traffic rule knowledge base is constructed, rule violation is identified, safety state is judged, and observation space feature vector is output;Three-stage memory retrieval mechanism is used, the most similar historical memory unit in memory bank is matched with current scene, if matching is successful, the corresponding multi-objective weight adjustment vector is extracted, otherwise, the observation space feature vector and user instruction are input into large language model, and multi-objective weight adjustment vector is generated based on large language model, and memory bank is updated;The weight in the comprehensive cost function is adjusted based on multi-objective weight adjustment vector, and the optimal path is generated based on the adjusted comprehensive cost function;The above steps are repeatedly executed at a fixed period to continuously update the optimal path.Compared with the prior art, the application has the advantages of higher safety, stronger applicability and the like.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving, and in particular to an adaptive autonomous driving decision-making method that integrates user preferences and traffic rules. Background Technology

[0002] With the continuous advancement of autonomous driving technology and its commercialization, the public is paying increasing attention to personalized driving experiences. When the human role shifts from driver to passenger, the ability to customize vehicle behavior, driving trajectory, and driving style according to individual preferences will significantly improve user satisfaction with autonomous vehicles. For example, a passenger might say, "I'm going to be late, please drive faster."

[0003] Chinese patent CN120066281A discloses a human-machine collaborative decision-making method for autonomous driving based on a fast and slow system, aiming to balance safety, flexibility, and controllability in autonomous driving scenarios. The model consists of a fast system based on deep reinforcement learning and a slow system based on a large language model. The fast system is responsible for real-time driving decisions and control, enabling rapid responses in short-term, high-frequency dynamic traffic environments. The slow system understands and parses human user instructions, combines environmental perception information to make high-level decisions and select target lanes, and then transmits this information to the fast system for execution. By introducing target lane and human instruction information into the observation space of the fast system and designing corresponding network structures and reward functions, the system can obey human instructions while ensuring safe driving, achieving "human-machine collaborative" autonomous driving. However, improving user experience cannot come at the expense of safety and compliance; these remain fundamental principles that autonomous driving systems must adhere to. This method only considers user instructions in the decision-making process to enhance the personalized driving experience, without comprehensively considering the impact of improved user experience on safety and compliance. Users' risk preferences vary significantly among different groups, making it impossible for autonomous vehicles to unconditionally cater to them completely. For example, aggressive risk preferences may lead to potentially dangerous driving behaviors such as speeding and cutting in front of others. Such behaviors not only increase the risk of collisions but also violate basic traffic regulations, thereby affecting public acceptance of autonomous vehicles.

[0004] Therefore, autonomous vehicles still face significant challenges in balancing personalized driving with adherence to traffic rules. This problem is essentially a multi-objective optimization problem, requiring real-time balancing of competing or conflicting objectives such as safety, operational efficiency, and passenger comfort across different driving scenarios. While this problem has attracted widespread attention in academia and various dynamic balancing methods have been proposed, most studies model traffic rule compliance and user preferences independently, lacking the ability to integrate them in real-time within a unified decision-making framework. Secondly, rule-based trajectory planning methods typically rely on fixed weights or manually set logic for multi-objective balancing, making them difficult to adapt to changing scenarios. Data-driven or generative methods, due to their "black box" nature, may produce illusions when directly applied to end-to-end decision outputs, lacking sufficient interpretability and credibility. Summary of the Invention

[0005] The purpose of this invention is to provide an adaptive autonomous driving decision-making method that integrates user preferences and traffic rules. This method is driven by a Large Language Model (LLM), which combines the advantages of LLM in integrating semantic user instructions and multi-regional traffic rule knowledge with the determinism and interpretability of rule-based trajectory planning methods.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] An adaptive autonomous driving decision-making method that integrates user preferences and traffic rules includes the following steps:

[0008] Step 1: Extract autonomous driving scenario information and user commands;

[0009] Step 2: Construct a traffic rule knowledge base, and identify violations based on autonomous driving scenario information and user instructions, determine the safety status, and output the observation space feature vector;

[0010] Step 3: Using a three-stage memory retrieval mechanism, the most similar historical memory unit to the current scene is matched in the memory bank. If the match is successful, the corresponding multi-objective weight adjustment vector is extracted; otherwise, proceed to step 4.

[0011] Step 4: Input the observed spatial feature vector and user instructions into the large language model, generate a multi-objective weight adjustment vector based on the large language model, and update the memory bank;

[0012] Step 5: Adjust the weights in the comprehensive cost function based on the multi-objective weight adjustment vector matched in Step 3 or generated in Step 4, and generate the optimal path based on the adjusted comprehensive cost function;

[0013] Step 6: Repeat steps 1-5 at a fixed period to continuously update the optimal path; during the loop, if a sudden high-risk event is detected or the user's instructions change significantly, jump to step 3.

[0014] Step 1 includes the following steps:

[0015] Step 1.1: Collect the current environmental information of the vehicle, including the vehicle kinematic parameters of the vehicle and surrounding traffic parameters, as well as the static environmental attributes in the traffic scene, and construct the scene state vector;

[0016] Step 1.2: Receive user voice commands through a voice device and convert them into text commands, or receive user commands through text input on the in-vehicle human-machine interface.

[0017] Step 1.1 includes the following steps:

[0018] Step 1.1.1: Obtain the kinematic parameters of the vehicle and construct the vehicle state vector s ego :

[0019] s ego =[x ego ,y ego ,v ego ,acc lat ,acc lon ,j ego ,p lat ,w ego ,l ego ]

[0020] Where, x ego ,y ego This represents the vehicle's position coordinates, v ego acc indicates the speed of the vehicle. lat ,acc lon Let j represent the lateral and longitudinal accelerations of the vehicle, respectively. ego p indicates the abruptness of the vehicle's movement. lat w represents the lateral offset of the vehicle relative to the reference path. ego ,l ego These represent the width and length of the vehicle, respectively.

[0021] Step 1.1.2: For each observable surrounding traffic participant Obtain its kinematic parameters and define its state vector s obs,i for:

[0022] s obs,i =[x i ,y i ,v i ,γ i,w i ,l i ]

[0023] in, Let x represent the currently observable set of surrounding traffic participants. i ,y i Let v be the location coordinates of the i-th traffic participant. i Let γ be the speed of the i-th traffic participant; i w is an indicator of the aggressiveness of the behavior of the i-th traffic participant. i ,l i Let represent the width and length of the i-th traffic participant, respectively;

[0024] The states of all surrounding traffic participants are combined into a set s of surrounding traffic object states. obs :

[0025]

[0026] Step 1.1.3: Extract static environmental attributes from the traffic scene and construct the environmental state vector s. env :

[0027] s env =[country,road type ,v limit ]

[0028] in, Indicates a country or region. For a collection of countries or regions, road type Indicates road type, v limtt This is the speed limit for the current road segment;

[0029] Step 1.1.4: Concatenate the vehicle state vector, the surrounding traffic object state set, and the environment state vector to construct a complete scene state vector s. sce Defined as:

[0030] s sce =[s env ,s ego ,s obs ].

[0031] Step 2 includes the following steps:

[0032] Step 2.1: Extract traffic regulations texts from multiple countries or regions, perform structured parsing of the traffic regulations texts from various countries, and construct a traffic rules knowledge base;

[0033] For each country or region k Its knowledge base structure is defined as follows:

[0034]

[0035] Among them, country k Represents country or region identifier; road type,k This indicates the road type classification for country or region k; Let T represent the set of traffic rule texts for country k. k,c This represents the text of the c-th rule of country k; V represents the violation type extracted from the text of rule c. k,c A set; This represents the penalty information P extracted from the text of rule c. k,c A set;

[0036] Step 2.2: Based on the scene state vector and user instructions extracted in Step 1, use cosine similarity to retrieve and identify violations in the traffic rule knowledge base, and output the violation state vector;

[0037] Step 2.2.1: For each violation type V k,c The semantic embedding vector ε is generated using a pre-trained language model. k,c =Emb(V k,c ), where Emb(.) represents the semantic embedding model;

[0038] Step 2.2.2: Obtain the current scene state vector s sce and user commands The two are concatenated and input into the same semantic embedding model Emb(.) to obtain the scene semantic vector ε. scene :

[0039]

[0040] Step 2.2.3: Calculate the scene semantic vector ε scene With each semantic embedding vector ε k,c Cosine similarity between them:

[0041]

[0042] Step 2.2.4: When the similarity exceeds the set threshold ∈1, it is determined that there is a potential violation, and the violation state vector is output:

[0043]

[0044] Among them, is vio A Boolean vector representing whether a violation exists, sim(ε) scene ,ε k,c When )>∈1, is vio=1, otherwise 0; This indicates a list of detected violations. This represents the corresponding set of penalty information.

[0045] Step 2.3: Output the security state vector according to the security rules of different countries;

[0046] Step 2.3.1: Extract security rules from the rule text of the knowledge base;

[0047] Step 2.3.2: Calculate and output the security state vector S based on the security rules. sft :

[0048] S sft =[d lon ,d lat ,ETTC,is safety ,is yield ]

[0049] Where, d lon ,d lat These represent the longitudinal and lateral relative distances, respectively; ETTC represents the extended collision time, calculated as follows: ETTC ego,i This represents the extended collision time between the vehicle's ego and surrounding traffic participants i. This represents the currently observable set of surrounding traffic participants; is safety A Boolean vector representing whether a security rule has been violated; is yield A Boolean vector indicating whether yielding to pedestrians or non-motorized vehicles was correct;

[0050] Step 2.4: Combine the scene state vector, violation state vector, and safe state vector with the current weight vector λ of the comprehensive cost function. d By piecing together elements, a unified structured observation space is constructed.

[0051]

[0052] Each memory unit m in the memory bank u Store partial scene state vectors s from historical scenes u Partial safety state vector User's natural language input or preference expression in this historical scenario And the multi-objective weight adjustment vector Δλ in this historical scenario u , represented as:

[0053]

[0054] Among them, s u ∈s sce Includes static environment attributes and some dynamic state information, defined as s u =[country u road type,u ,v limit,u ,v ego,u ,γ u ], where v limit,u Indicates the road speed limit in the current historical scenario, v ego,u γ represents the vehicle speed in the current historical scene. u This indicates the level of aggression among surrounding traffic participants in the current historical context; Includes partial security status information, represented as: Among them, ETTC u d represents the extended collision time of the current historical scene. lon,u and d lat,u These represent the vertical distance and the horizontal distance, respectively; additionally, the multi-objective weight adjustment vector is also stored during each iteration, denoted as Δλ. u =[Δλ safety ,Δλ efficiency ,Δλ comfort_lon ,Δλ comfort_lat ,Δλ tracking ], where Δλ safety ,Δλ efficiency ,Δλ comfort_lon ,Δλ comfort_lat ,Δλ tracking These represent the weight adjustments for the safety cost function, efficiency cost function, longitudinal comfort cost function, lateral comfort cost function, and path deviation cost function, respectively.

[0055] Step 3 includes the following steps:

[0056] Step 3.1: Based on the static environment information of the current scene, retrieve from the memory... Select a set of candidate memory units that match the static environment. As shown in the following formula:

[0057]

[0058] Where, m u Represents a memory unit, country u road type,u These represent the country and road type information stored in the memory unit, respectively. d road type,d These represent the country and road type of the current scene, respectively;

[0059] Step 3.2: For the candidate memory unit set Numerical features of the current scene are calculated using the weighted Manhattan distance method. Features of historical scenes The cosine similarity is shown in the following formula:

[0060]

[0061] Where, ω f The weight of the f-th numerical feature is represented by s, where F represents the dimension of the numerical feature. d s u These represent partial scene state vectors of the current scene and historical scenes in the memory bank, respectively. These represent partial security state vectors of the current scene and the historical scene in the memory bank, respectively.

[0062] If the calculated cosine similarity distance is less than the preset threshold ∈ 2, then the memory unit is retained, forming the set of memory units after numerical similarity filtering.

[0063]

[0064] Step 3.3: For the set Each memory unit in the memory measures the current user instruction input. With historical instructions The semantic similarity between them is calculated using the vector representation function φ(·) and cosine similarity:

[0065]

[0066] Step 3.4: Determine the most relevant memory units based on the semantic similarity calculation results in Step 3.3.

[0067]

[0068] Step 3.5: If the most relevant memory unit is successfully retrieved Then extract the corresponding multi-objective weight adjustment vector. Proceed directly to step 5; if the most relevant memory unit is not found. Then proceed to step 4.

[0069] Step 4 includes the following steps:

[0070] Step 4.1: Obtain the observed spatial feature vector Transform it into a form that conforms to the rules of natural language, and combine it with the current user command. Input them together into the large language model;

[0071] Step 4.2: Based on the input content, the large language model, through its scene understanding and semantic reasoning capabilities, outputs a natural language response text R. b :

[0072]

[0073] Define a parsing function Parse(·) to parse the natural language response text R. b Extract the multi-objective weight adjustment vector Δλ b :

[0074] Δλ b =Parse(R b )

[0075] Where, Δλ b ={Δλ safety ,Δλ efficiency ,Δλ comfort_lon ,Δλ comfort_lat ,Δλ tracking}, Δλ safety ,Δλ efficiency ,Δλ comfort_lon ,Δλ comfort_lat ,Δλ tracking These represent the weight adjustments for the safety cost function, efficiency cost function, longitudinal comfort cost function, lateral comfort cost function, and path deviation cost function, respectively.

[0076] Step 4.3: Store the partial scene description vector, partial security state vector, user natural language input or preference expression in the scene, and the extracted multi-objective weight adjustment vector into the memory unit of the memory bank.

[0077] The natural language response text comprehensively considers the task objectives, security status, violation status, and user preference factors of the current scenario, and adopts a four-stage chain reasoning process to output the results, which includes violation and security assessment, weight and user preference assessment, security-efficiency-comfort-path deviation balance, and constraint and compliance checks.

[0078] In step 5, the weights in the comprehensive cost function are adjusted based on the multi-objective weight adjustment vector matched in step 3 or generated in step 4, specifically as follows:

[0079] Based on the current weight vector λ corresponding to the d-th objective in the comprehensive cost function d With the multi-objective weight adjustment vector Δλ d The weights of each objective are normalized and updated to ensure that the sum of the weights of all objectives remains 1, reflecting their relative importance. The update formula is as follows:

[0080]

[0081] Wherein, the multi-objective weight adjustment vector Δλ d Adjust the multi-objective weight vector Δλ matched in step 3. u *or the multi-objective weight adjustment vector Δλ generated in step 4 b The d-th weight adjustment value in the data. The adjusted weight value for the d-th objective.

[0082] In step 5, generating the optimal path based on the adjusted comprehensive cost function specifically involves:

[0083] A sampling-based path planning method is used to calculate the comprehensive cost function value of each candidate path. This comprehensive cost function considers five key objectives: safety, efficiency, longitudinal comfort, lateral comfort, and path deviation.

[0084]

[0085] Where Cost(n) represents the comprehensive cost function value of the nth candidate path; These are the adjusted normalized weights corresponding to each objective function;

[0086] C safe (n) is the safety cost function of the nth path, defined as the collision risk with surrounding traffic participants. This collision risk is quantified and evaluated using Mahalanobis distance. The probability p of the nth path colliding with surrounding traffic participant i at time t is... c (n,t,i) is defined as:

[0087]

[0088] in, Let δ represent the square of the Mahalanobis distance between the vehicle and surrounding traffic participant i at time t, where δ is a constant used for numerical stability.

[0089] The safety cost function of the nth path is expressed as the cost of all observable surrounding traffic participants within the planning period T. The result of summing and normalizing the collision probabilities:

[0090]

[0091] Where norm(·) represents the normalization function;

[0092] C efficiency (n) is the efficiency cost function of the nth path, expressed as the average speed versus the speed limit V. limit The normalized result of the bias is defined as:

[0093]

[0094] Where v(n,t) represents the velocity of the nth path at time t;

[0095] C comfort_lat (n), C comfort_lon (n) represents the lateral and longitudinal comfort cost functions for the nth path, respectively, expressed by the acceleration change rate jerk (Jerk), and defined as the normalized result of the cumulative Jerk values ​​in the lateral and longitudinal directions:

[0096]

[0097] Among them, Jerk lat (n,t), Jerk lon (n,t) represent the horizontal and vertical Jerk values ​​of the nth path at time t, respectively;

[0098] C path (n) is the path deviation cost function for the nth path, which is expressed as the average weighted lateral deviation between the vehicle's path and the reference path:

[0099]

[0100] Where d represents the set of lateral offsets of the path, d end Let len(d) be the lateral offset of the endpoint of the nth path, and let len(d) represent the number of sampling points contained in the path, i.e., the path length.

[0101] Multiple candidate path samples are generated within each planning cycle, based on the updated weight parameters. Calculate the overall cost function Cost(n) for each path and select the path with the lowest cost as the optimal path.

[0102] Compared with the prior art, the present invention has the following beneficial effects:

[0103] (1) Inspired by the Retrieval-Augmented Generation (RAG) model paradigm, this invention first retrieves relevant rules from a structured traffic rule knowledge base, extracts key features such as environmental state, violation state, and safety state, and encodes them as prompt information input to the LLM to support context-based multi-objective weight dynamic adjustment reasoning, thereby guiding rule-based trajectory planning. This method significantly enhances the model's ability to adapt to different traffic rules.

[0104] (2) In order to improve the generalization ability of unknown scenarios and reduce redundant reasoning, the present invention integrates a memory module to store historical planning experience and retrieve similar cases to assist the current decision-making.

[0105] (3) This invention realizes the real-time balance between users’ personalized driving preferences and compliance with traffic rules in autonomous vehicles, which has the characteristics of higher safety and can significantly improve the user experience. Attached Figure Description

[0106] Figure 1 This is a flowchart of the method of the present invention.

[0107] Figure 2 This is a diagram of the chain-like reasoning process. Detailed Implementation

[0108] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0109] This embodiment provides an adaptive autonomous driving decision-making method that integrates user preferences and traffic rules, such as Figure 1 As shown, it includes the following steps:

[0110] Step 1: Extract autonomous driving scenario information and user commands.

[0111] Step 1.1: Collect environmental information about the vehicle's current location, including but not limited to kinematic parameters such as vehicle position, speed, and acceleration of the vehicle and surrounding traffic parameters, as well as static environmental attributes such as road type and speed limit in the traffic scene, and construct a scene state vector.

[0112] Step 1.1.1: Obtain the kinematic parameters of the vehicle and construct the vehicle state vector s ego :

[0113] s ego =[x ego ,y ego ,v ego ,acc lat ,acc lon ,j ego ,p lat ,w ego ,l ego ]

[0114] Where, x ego ,y ego This represents the vehicle's position coordinates, v ego acc indicates the speed of the vehicle. lat ,acc lon Let j represent the lateral and longitudinal accelerations of the vehicle, respectively. ego The jerk (p) indicates the vehicle's agility. lat w represents the lateral offset of the vehicle relative to the reference path.ego ,l ego These represent the width and length of the vehicle, respectively.

[0115] Step 1.1.2: For each observable surrounding traffic participant Obtain its kinematic parameters and define its state vector s obs,i for:

[0116] s obs,i =[x i ,y i ,v i ,γ i ,w i ,l i ]

[0117] in, Let x represent the currently observable set of surrounding traffic participants. i ,y i Let v be the location coordinates of the i-th traffic participant. i Let γ be the speed of the i-th traffic participant; i w represents the behavioral aggression index of the i-th traffic participant; a smaller value indicates more aggressive behavior. i ,l i Let represent the width and length of the i-th traffic participant, respectively.

[0118] The states of all surrounding traffic participants are combined into a set s of surrounding traffic object states. obs :

[0119]

[0120] Step 1.1.3: Extract static environmental attributes from the traffic scene and construct the environmental state vector s. env :

[0121] s env =[country,road type ,v limit ]

[0122] in, Indicates a country or region. For a collection of countries or regions, road type Indicates the type of road, such as highway, urban road, or rural road, v limit This is the speed limit for the current road segment.

[0123] Step 1.1.4: Concatenate the vehicle state vector, the surrounding traffic object state set, and the environment state vector to construct a complete scene state vector s. sce Defined as:

[0124] s sce =[s env ,s ego ,s obs ]

[0125] Step 1.2: Receive user voice commands via a voice device and convert them into text format. Alternatively, user commands can be received via text input through the in-vehicle human-machine interface. For example, user commands could be such as "arrive as soon as possible" or "drive more steadily".

[0126] Step 2: Construct a traffic rule knowledge base, and identify violations based on autonomous driving scenario information and user instructions, determine the safety status, and output the observation space feature vector.

[0127] Step 2.1: Extract traffic regulations texts from multiple countries or regions, perform structured parsing of the traffic regulations texts from various countries, and construct a traffic rules knowledge base.

[0128] For each country or region k Its knowledge base structure is defined as follows:

[0129]

[0130] Among them, country k Represents country or region identifier; road type,k This indicates the classification of road types in country or region k, such as urban roads, highways, rural roads, etc. Let T represent the set of traffic rule texts for country k. k,c This represents the text of the c-th rule of country k; V represents the violation type extracted from the text of rule c. k,c A set of, for example, speeding, failing to yield; This represents the penalty information P extracted from the text of rule c. k,c A collection of penalties, such as fines, demerit points, and imprisonment;

[0131] Step 2.2: Based on the scene state vector s extracted in Step 1 sce and user instructions Cosine similarity is used to retrieve and identify traffic violations in a traffic rule knowledge base, and a violation status vector is output.

[0132] Step 2.2.1: For each violation type V k,c The semantic embedding vectors are generated using a pre-trained language model (such as all-MiniLM-L6-v2). Where Emb(.) represents the semantic embedding model;

[0133] Step 2.2.2: Obtain the current scene state vector s sce and user commands The two are concatenated and input into the same semantic embedding model Emb(.) to obtain the scene semantic vector ε. scene :

[0134]

[0135] Step 2.2.3: Calculate the scene semantic vector ε scene With each semantic embedding vector ε k,c Cosine similarity between them:

[0136]

[0137] Step 2.2.4: When the similarity exceeds the set threshold ∈1, it is determined that there is a potential violation, and the violation state vector is output:

[0138]

[0139] Among them, is vio A Boolean vector representing whether a violation exists, sim(ε) scene ,ε k,c When )>∈1, is vio =1, otherwise 0; This indicates a list of detected violations. This represents the corresponding set of penalty information.

[0140] Step 2.3: Output the security state vector according to the security rules of different countries.

[0141] Step 2.3.1: Refer to the reference "Digitizing traffic rules to guide automated vehicle trajectory planning" to extract safety rules from the rule text in the knowledge base. For example, China uses the "three-second rule" for safety distance assessment, and Germany stipulates that the safety distance is half the vehicle speed. Extract relevant rule information.

[0142] Step 2.3.2: Calculate and output the security state vector S based on the security rules. sft :

[0143] S sft =[d lon ,d lat ,ETTC,is safety ,is yield ]

[0144] Where, d lon ,d lat These represent the longitudinal and lateral relative distances, respectively; ETTC represents the extended collision time, calculated as follows: ETTC ego,i This represents the extended collision time between the vehicle's ego and surrounding traffic participants i. This represents the currently observable set of surrounding traffic participants; is safety A Boolean vector representing whether a safety rule has been violated, determined according to the safety rules extracted in step 2.3.1. For example, if in Germany, the longitudinal distance is less than half the vehicle speed, then is... safety =0 indicates unsafe; is yield A Boolean vector indicating whether yielding to pedestrians or non-motorized vehicles is correct.

[0145] The calculation of extended collision time is referenced in the literature "Conflict Extraction and Characteristics Analysis at Signalized Intersections Using Trajectory Data", which will not be repeated here in this embodiment.

[0146] Step 2.4: Combine the scene state vector, violation state vector, and safe state vector with the current weight vector λ of the comprehensive cost function. d By piecing together elements, a unified structured observation space is constructed.

[0147]

[0148] Step 3: Using a three-stage memory retrieval mechanism, the most similar historical memory unit to the current scene is matched in the memory bank. If the match is successful, the corresponding multi-objective weight adjustment vector is extracted; otherwise, proceed to step 4.

[0149] Step 3.1: Based on the static environment information of the current scene, retrieve from the memory... Select a set of candidate memory units that match the static environment. As shown in the following formula:

[0150]

[0151] Where, m u Represents a memory unit, country u road type,u These represent the country and road type information stored in the memory unit, respectively. d road type,d These represent the country and road type of the current scene, respectively.

[0152] Each memory unit m in the memory bank u Store partial scene state vectors s from historical scenes u Partial safety state vector User's natural language input or preference expression in this historical scenario And the multi-objective weight adjustment vector Δλ in this historical scenario u , represented as:

[0153]

[0154] Among them, s u ∈s sce Includes static environment attributes and some dynamic state information, defined as s u =[country u road type,u ,v limit,u ,v ego,u ,γ u ], where v limit,u Indicates the road speed limit in the current historical scenario, v ego,u γ represents the vehicle speed in the current historical scene. u This indicates the level of aggression among surrounding traffic participants in the current historical context; Includes partial security status information, represented as: Among them, ETTC u d represents the extended collision time of the current historical scene. lon,u and d lat,u These represent the vertical distance and the horizontal distance, respectively; additionally, the multi-objective weight adjustment vector is also stored during each iteration, denoted as Δλ. u =[Δλ safety ,Δλ efficiency ,Δλ comfort_lon ,Δλ comfort_lat ,Δλ tracking ], where Δλ safety ,Δλ efficiency ,Δλ comfort_lon ,Δλ comfort_lat ,Δλ tracking These represent the weight adjustments for the safety cost function, efficiency cost function, longitudinal comfort cost function, lateral comfort cost function, and path deviation cost function, respectively.

[0155] In this embodiment, the calculation method of radicality is based on pages 213–222 of the book "Traffic Flow Theory: Characteristics, Experimental Methods, and Numerical Techniques", and will not be repeated here.

[0156] Step 3.2: For the candidate memory unit set Numerical features of the current scene are calculated using the weighted Manhattan distance method. Features of historical scenes The cosine similarity is shown in the following formula:

[0157]

[0158] Where, ω f The weight of the f-th numerical feature is represented by s, where F represents the dimension of the numerical feature. d s u These represent partial scene state vectors of the current scene and historical scenes in the memory bank, respectively. These represent partial security state vectors of the current scene and the historical scene in the memory bank, respectively; s d =[country d road type,d ,v limit,d ,v ego,d ,γ d ], The subscript d represents the current scene, and the subscript u represents the historical scene.

[0159] If the calculated cosine similarity distance is less than the preset threshold ∈ 2, then the memory unit is retained, forming the set of memory units after numerical similarity filtering.

[0160]

[0161] Step 3.3: For the set Each memory unit in the memory measures the current user instruction input. With historical instructions The semantic similarity between them is calculated using the vector representation function φ(·) and cosine similarity:

[0162]

[0163] Step 3.4: Determine the most relevant memory units based on the semantic similarity calculation results in Step 3.3.

[0164]

[0165] Step 3.5: If the most relevant memory unit is successfully retrieved Then extract the corresponding multi-objective weight adjustment vector. Proceed directly to step 5; if the most relevant memory unit is not found. Then proceed to step 4.

[0166] Step 4: Input the observed spatial feature vector and user instructions into the large language model, generate a multi-objective weight adjustment vector based on the large language model, and update the memory bank.

[0167] Step 4.1: Obtain the observed spatial feature vector Transform it into a form that conforms to the rules of natural language, and combine it with the current user command. They are all input into the large language model.

[0168] Step 4.2: Based on the input content, the large language model, through its scene understanding and semantic reasoning capabilities, outputs a natural language response text R. b :

[0169]

[0170] The natural language response text comprehensively considers the task objectives, security status, violation status, and user preference factors of the current scenario. It adopts a four-stage chain reasoning process to output results, which includes violation and security assessment, weight and user preference assessment, security-efficiency-comfort-path deviation balance, and constraint and compliance checks, in order to improve the interpretability of the model.

[0171] Meanwhile, to improve the reliability of policy generation and maintain compliance with security and consistency constraints, a set of adaptive adjustment rules and coverage mechanisms are introduced:

[0172] (1) Adaptive weight adjustment rule

[0173] The target weight is dynamically adjusted based on detected anomalies or specific condition thresholds:

[0174] ① Safety Critical Conditions: When the Extended Time to Collision (ETTC) is less than 1.0 second, the vehicle jerk exceeds 5.0 m / s. 3 or longitudinal acceleration exceeding 2.4 m / s² 2 At the same time, the weights related to safety or comfort are increased respectively to encourage vehicles to adopt more conservative risk-avoidance behaviors;

[0175] ② Path deviation control: When the lateral deviation exceeds a preset threshold, the path deviation weight is increased to promote trajectory correction and lane keeping;

[0176] ③ Adjustment range constraints: All weight updates are strictly limited to the range of [-1.0, 1.0] to avoid sudden or unstable strategy changes.

[0177] (2) Rule-based coverage mechanism

[0178] In specific high-risk scenarios, default targets or user-defined preferences can be overridden:

[0179] ① Violation-based coverage: When obvious traffic violations (such as dangerous driving, failure to yield, etc.) are detected, passenger preference weights are suppressed and safety constraints are given priority;

[0180] ② Emergency handling: In extremely dangerous situations, some traffic rules may be temporarily relaxed to prioritize collision avoidance and risk minimization.

[0181] Specifically, such as Figure 2 As shown, a chain-of-action (CoT) reasoning process is employed, with each stage progressively refining the decision based on the results of the previous stage:

[0182] Phase 1: Violation and Security Assessment

[0183] The initial phase assesses the current driving status to identify traffic violations and determine key risks. This process includes analyzing for behaviors such as failure to yield to vulnerable road users and speeding, and monitoring safety surrogate indicators such as ETTC (Electronic Toll Collection). By comparing these indicators to preset safety thresholds, the vehicle's risk level is determined, and key constraints for subsequent decisions are defined. For example, if a vehicle is detected speeding and the ETTC is 0.7 seconds, below the critical value of 1.0 second, the scenario is deemed unsafe, and any conflicting user preferences (such as acceleration requests) are overridden.

[0184] Phase Two: Weighting and User Preference Assessment

[0185] Based on the state determined in Phase 1, this phase evaluates whether the current weight configuration is consistent with the evaluation results and determines the extent to which user preferences can be satisfied without violating safety or traffic rules.

[0186] Phase Three: Generation of Four-Dimensional Balance Strategy

[0187] In this stage, a weight adjustment strategy is generated based on the "four-dimensional balance" principle, with safety, efficiency, comfort, and path deviation as the main decision objectives. By dynamically adjusting the weights of these competing objectives, risks are reduced and overall driving performance is improved. All adjustments are confined to the range [-1.0, 1.0].

[0188] Phase Four: Constraints and Compliance Checks

[0189] The final stage involves a comprehensive verification of the proposed weight adjustment scheme to ensure that all weight values ​​are within the allowable range and to confirm that the generated strategy complies with applicable traffic rules and system restrictions, thereby ensuring its legality and feasibility before deployment.

[0190] Define a parsing function Parse(·) to parse the natural language response text R. b Extract the multi-objective weight adjustment vector Δλ b :

[0191] Δλ b =Parse(R b )

[0192] Where, Δλ b ={Δλ safety ,Δλ efficiency ,Δλ comfort_lon ,Δλ comfort_lat ,Δλ tracking}

[0193] Step 4.3: To avoid information redundancy and improve subsequent retrieval efficiency, the memory module selectively stores a portion of the observation data and the corresponding strategy weight adjustment results. This includes storing a portion of the scenario description vector, a portion of the security state vector, the user's natural language input or preference expression in this scenario, and the extracted multi-objective weight adjustment vector into the memory bank. The memory unit in the middle.

[0194] Step 5: Adjust the weights in the comprehensive cost function based on the multi-objective weight adjustment vector matched in Step 3 or generated in Step 4, and generate the optimal path based on the adjusted comprehensive cost function;

[0195] Step 5.1: Adjust the weights in the comprehensive cost function based on the multi-objective weight adjustment vector matched in Step 3 or generated in Step 4.

[0196] Based on the current weight vector λ corresponding to the d-th objective in the comprehensive cost function d With the multi-objective weight adjustment vector Δλ d The weights of each objective are normalized and updated to ensure that the sum of the weights of all objectives remains 1, reflecting their relative importance. The update formula is as follows:

[0197]

[0198] Wherein, the multi-objective weight adjustment vector Δλ d Adjust the multi-objective weight vector Δλ matched in step 3. u *or the multi-objective weight adjustment vector Δλ generated in step 4 b The d-th weight adjustment value in the data. The adjusted weight value for the d-th objective.

[0199] Step 5.2: Generate the optimal path based on the adjusted comprehensive cost function.

[0200] User preferences (such as smooth driving and increased speed) and compliance goals (such as speed limits, lane keeping, and priority rules) belong to high-level semantic goals. Therefore, they are not directly used as optimization objectives. Instead, they are modeled by constructing a measurable low-level cost function, and a weight adjustment mechanism is used to map high-level semantic vectors to low-level behaviors. A sampling-based path planning method is used to calculate the comprehensive cost function value of each candidate path. The comprehensive cost function considers five key objectives, including safety, efficiency, longitudinal comfort, lateral comfort, and path deviation.

[0201]

[0202] Where Cost(n) represents the comprehensive cost function value of the nth candidate path; These are the adjusted normalized weights corresponding to each objective function.

[0203] C safe (n) is the safety cost function of the nth path, defined as the collision risk with surrounding traffic participants. This collision risk is quantified and evaluated using Mahalanobis distance. The probability p of the nth path colliding with surrounding traffic participant i at time t is... c (n,t,i) is defined as:

[0204]

[0205] in, Let δ represent the square of the Mahalanobis distance between the vehicle and surrounding traffic participant i at time t, where δ is a constant used for numerical stability.

[0206] The safety cost function of the nth path is expressed as the cost of all observable surrounding traffic participants within the planning period T. The result of summing and normalizing the collision probabilities:

[0207]

[0208] Here, norm(·) represents the normalization function.

[0209] C efficiency (n) is the efficiency cost function of the nth path, expressed as the average speed versus the speed limit V. limit The normalized result of the bias is defined as:

[0210]

[0211] Where v(n,t) represents the velocity of the nth path at time t.

[0212] C co,fort_lat (n), C comfort_lon (n) represents the lateral and longitudinal comfort cost functions for the nth path, respectively, denoted by the jerk factor (Jerk value) of the acceleration change rate. A smaller Jerk value indicates smoother vehicle control and a better ride experience. The value is defined as the normalized result of the cumulative Jerk values ​​in the lateral and longitudinal directions.

[0213]

[0214] Among them, Jerk lat (n,t), Jerk lon (n,t) represent the horizontal and vertical Jerk values ​​of the nth path at time t, respectively.

[0215] C path (n) is the path deviation cost function for the nth path, which is expressed as the average weighted lateral deviation between the vehicle's path and the reference path:

[0216]

[0217] Where d represents the set of lateral offsets along the path (in the Frenet coordinate system, the distance between the vehicle and the reference path in the normal direction), d end Let len(d) be the lateral offset of the endpoint of the nth path, and let len(d) represent the number of sampling points contained in the path, i.e., the path length.

[0218] Multiple candidate path samples are generated within each planning cycle, based on the updated weight parameters. Calculate the overall cost function Cost(n) for each path and select the path with the lowest cost as the optimal path.

[0219] Step 6: Repeat steps 1-5 at a fixed period (e.g., every 2 seconds) to continuously update the optimal path; during the loop, if a sudden high-risk event is detected (e.g., an obstacle suddenly appears) or the user command changes significantly (e.g., a priority command is changed), jump to step 3.

[0220] To verify the effectiveness of the method of this invention, the CommonRoad dataset was used for testing. This dataset provides standardized traffic scenarios for autonomous driving trajectory planning and decision-making research, and is broad and representative. In the experiment, 1,900 typical scenarios were selected from the dataset as test samples. The experimental environment included an RTX 3060 GPU and 13th Gen 1300 GPU. Core TMi5-13400F×16 CPU to ensure computing performance.

[0221] To facilitate the expansion of memory units, the dataset was divided into two subsets: 80% of the scenarios were used in the training phase to gradually build the memory bank. The remaining 20% ​​of the scenarios were used for testing to evaluate the adaptability of the invention under unknown conditions. During the testing process, to simulate the reaction time of a human driver, the user preference instructions and inference module were periodically activated every 2 seconds.

[0222] To verify the optimal model selection of the method of this invention, the reasoning and decision-making capabilities of various Large Language Models (LLMs) were compared and evaluated. Candidate models included Qwen-Turbo-Latest, DeepSeek-V3, DeepSeek-R1, and Llama-4-Scout-17B-16E-Instruct. The average cumulative collision probability (p0.05) was compared. mean Based on the analysis of efficiency and single-step inference time, Qwen-Turbo-Latest was found to achieve the best balance between efficiency and performance, and was therefore selected as the base model for the high-level inference and decision-making module (see Table 1).

[0223] Table 1 Performance Comparison of Different LLMs

[0224]

[0225] In addition, during the testing phase, comparisons were made with the following baseline trajectory planning:

[0226] 1. Freneix Planner: A traditional rule-based approach that uses fixed-cost weights.

[0227] 2. Freneix+QL: An improved version of Freneix that uses reinforcement learning to dynamically adjust cost weights.

[0228] 3. Rule Planner: A planner that incorporates traffic rule constraints, but still relies on fixed weights.

[0229] The method's performance is evaluated using the following six key metrics:

[0230] 1. Success rate: The percentage of scenarios that run successfully.

[0231] 2. Average planning time (t) mean ): The average time to generate the optimal trajectory in a single step.

[0232] 3. Average cumulative collision probability (p) mean ): Average collision probability.

[0233] 4. Average velocity (v) mean ): The average speed of trajectory planning.

[0234] 5. Average rate of change of acceleration (j mean ): The average value of jerk (a measure of urgency).

[0235] 6. Average lateral offset (d) mean ): The average distance between the vehicle and the reference path in the normal direction.

[0236] As shown in Table 2, the method proposed in this invention exhibits significant advantages over the baseline method, with a success rate increase of 5.0% to 12.1%, a planning time reduction of 8.3% to 85.5%, and an average cumulative collision probability reduction of 72.8% to 95.0%. The results demonstrate that the method proposed in this invention not only significantly improves the safety of trajectory planning and effectively transforms potential collision events into controllable risks in dynamic and uncertain environments, but also optimizes decision paths without sacrificing operational efficiency, providing reliable support for the development of autonomous driving technology.

[0237] Table 2 Comparison of the present invention with mainstream trajectory planning methods

[0238] Freneix Planner 69.74％ 0.74 7.01 6.44 1.72 0.21 Rule Planner 70.00％ 1.26 6.59 6.40 2.21 0.21 Freneix+QL 74.47％ 0.20 1.29 6.76 0.86 0.32 This invention 78.16％ 0.18 0.35 6.74 0.87 0.32

[0239] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. An adaptive autonomous driving decision-making method that integrates user preferences and traffic rules, characterized in that, Includes the following steps: Step 1: Extract autonomous driving scenario information and user commands; Step 2: Construct a traffic rule knowledge base, and identify violations based on autonomous driving scenario information and user instructions, determine the safety status, and output the observation space feature vector; Step 3: Using a three-stage memory retrieval mechanism, the most similar historical memory unit to the current scene is matched in the memory bank. If the match is successful, the corresponding multi-objective weight adjustment vector is extracted; otherwise, proceed to step 4. Step 4: Input the observed spatial feature vector and user instructions into the large language model, generate a multi-objective weight adjustment vector based on the large language model, and update the memory bank; Step 5: Adjust the weights in the comprehensive cost function based on the multi-objective weight adjustment vector matched in Step 3 or generated in Step 4, and generate the optimal path based on the adjusted comprehensive cost function; Step 6: Repeat steps 1-5 at a fixed period to continuously update the optimal path; during the loop, if a sudden high-risk event is detected or the user's instructions change significantly, jump to step 3. Step 3 includes the following steps: Step 3.1: Based on the static environment information of the current scene, retrieve from the memory... Select a set of candidate memory units that match the static environment. As shown in the following formula: in, Represents a memory unit. , These represent the country and road type information stored in the memory unit, respectively. , These represent the country and road type of the current scene, respectively; Step 3.2: For the candidate memory unit set The weighted Manhattan distance method is used to calculate the numerical features of the current scene. Features of historical scenes The cosine similarity is shown in the following formula: in, Indicates the first Weights of numerical features, Represents the dimension of numerical features. , These represent partial scene state vectors of the current scene and historical scenes in the memory bank, respectively. , These represent partial security state vectors of the current scene and the historical scene in the memory bank, respectively. If the calculated cosine similarity distance is less than the preset threshold If the memory cell is selected, it will be retained, forming a set of memory cells after numerical similarity filtering. : Step 3.3: For the set Each memory unit in the memory measures the current user instruction input. With historical instructions The semantic similarity between them is represented by a vector function. Similarity calculation: Step 3.4: Determine the most relevant memory units based on the semantic similarity calculation results in Step 3.

3. : Step 3.5: If the most relevant memory unit is successfully retrieved Then extract the corresponding multi-objective weight adjustment vector. If no most relevant memory unit is found, proceed directly to step 5; Then proceed to step 4.

2. The adaptive autonomous driving decision-making method integrating user preferences and traffic rules according to claim 1, characterized in that, Step 1 includes the following steps: Step 1.1: Collect the current environmental information of the vehicle, including the vehicle kinematic parameters of the vehicle and surrounding traffic parameters, as well as the static environmental attributes in the traffic scene, and construct the scene state vector; Step 1.2: Receive user voice commands through a voice device and convert them into text commands, or receive user commands through text input on the in-vehicle human-machine interface.

3. The adaptive autonomous driving decision-making method integrating user preferences and traffic rules according to claim 2, characterized in that, Step 1.1 includes the following steps: Step 1.1.1: Obtain the kinematic parameters of the vehicle and construct the vehicle state vector. : in, Indicates the vehicle's position coordinates. Indicates the speed of the vehicle. These represent the lateral and longitudinal accelerations of the vehicle, respectively. Indicates the speed of movement of the vehicle. This represents the lateral offset of the vehicle relative to the reference path. These represent the width and length of the vehicle, respectively. Step 1.1.2: For each observable surrounding traffic participant Obtain its kinematic parameters and define its state vector. for: in, This represents the currently observable set of surrounding traffic participants. For the first The location coordinates of each traffic participant For the first The speed of each traffic participant; For the first An indicator of the aggressiveness of traffic participants' behavior; They represent the first Width and length of each traffic participant; Combine the states of all surrounding traffic participants into a set of surrounding traffic object states. : Step 1.1.3: Extract static environmental attributes from the traffic scene and construct an environmental state vector. : in, Indicates a country or region. A collection of countries or regions Indicates the road type. This is the speed limit for the current road segment; Step 1.1.4: Concatenate the vehicle state vector, the surrounding traffic object state set, and the environment state vector to construct a complete scene state vector. Defined as: 。 4. The adaptive autonomous driving decision-making method integrating user preferences and traffic rules according to claim 1, characterized in that, Step 2 includes the following steps: Step 2.1: Extract traffic regulations texts from multiple countries or regions, perform structured parsing of the traffic regulations texts from various countries, and construct a traffic rules knowledge base; For each country or region Its knowledge base structure is defined as follows: in, represent Country or region identifier; express National or regional road type classification; A collection of traffic rules texts from various countries. express country's first Rule text; Indicates from the first Types of violations extracted from the rule text A set; Indicates from the first Penalty information extracted from the rule text A set; Step 2.2: Based on the scene state vector and user instructions extracted in Step 1, use cosine similarity to retrieve and identify violations in the traffic rule knowledge base, and output the violation state vector; Step 2.2.1: For each type of violation The semantic embedding vectors are generated using a pre-trained language model. ,in, Represents a semantic embedding model; Step 2.2.2: Obtain the current scene state vector and user commands The two are concatenated and then input into the same semantic embedding model. To obtain the scene semantic vector : Step 2.2.3: Calculate the scene semantic vector With each semantic embedding vector Cosine similarity between them: Step 2.2.4: When the similarity exceeds the set threshold When a potential violation is detected, a violation state vector is output: in, A Boolean vector representing whether a violation exists. hour, =1, otherwise 0; This indicates a list of detected violations. ; This represents the corresponding set of penalty information. ; Step 2.3: Output the security state vector according to the security rules of different countries; Step 2.3.1: Extract security rules from the rule text of the knowledge base; Step 2.3.2: Calculate and output the security state vector based on the security rules. : in, These represent the relative distances in the longitudinal and transverse directions, respectively. This represents the extended collision time, calculated as follows: , Indicates a bicycle With surrounding traffic participants Extended collision time, This represents the set of currently observable surrounding traffic participants; A Boolean vector representing whether a security rule has been violated; A Boolean vector indicating whether yielding to pedestrians or non-motorized vehicles was correct; Step 2.4: Combine the scene state vector, violation state vector, and safe state vector with the current weight vector of the comprehensive cost function. By piecing together elements, a unified structured observation space is constructed. : 。 5. The adaptive autonomous driving decision-making method integrating user preferences and traffic rules according to claim 1, characterized in that, Each memory unit in the memory bank Store partial scene state vectors from historical scenes Partial safety state vector User's natural language input or preference expression in this historical scenario and the multi-objective weight adjustment vector in this historical scenario , represented as: in, Includes static environment attributes and some dynamic state information, defined as ,in, This indicates the road speed limit in the current historical context. This indicates the vehicle's speed in the current historical scenario. This indicates the level of aggression among surrounding traffic participants in the current historical context; Includes partial security status information, represented as: ,in, Indicates the extended collision time of the current historical scene. and These represent the vertical distance and the horizontal distance, respectively; additionally, the multi-objective weight adjustment vector is also stored during each iteration, represented as... ,in, These represent the weight adjustments for the safety cost function, efficiency cost function, longitudinal comfort cost function, lateral comfort cost function, and path deviation cost function, respectively.

6. The adaptive autonomous driving decision-making method integrating user preferences and traffic rules according to claim 1, characterized in that, Step 4 includes the following steps: Step 4.1: Obtain the observed spatial feature vector Transform it into a form that conforms to the rules of natural language, and combine it with the current user command. Input them together into the large language model; Step 4.2: Based on the input content, the large language model, through its scene understanding and semantic reasoning capabilities, outputs natural language response text. : Define analytic function From natural language response text Extracting multi-objective weight adjustment vectors : in, , These represent the weight adjustments for the safety cost function, efficiency cost function, longitudinal comfort cost function, lateral comfort cost function, and path deviation cost function, respectively. Step 4.3: Store the partial scene description vector, partial security state vector, user natural language input or preference expression in the scene, and the extracted multi-objective weight adjustment vector into the memory unit of the memory bank.

7. The adaptive autonomous driving decision-making method integrating user preferences and traffic rules according to claim 1, characterized in that, The natural language response text comprehensively considers the task objectives, security status, violation status, and user preference factors of the current scenario, and adopts a four-stage chain reasoning process to output the results, which includes violation and security assessment, weight and user preference assessment, security-efficiency-comfort-path deviation balance, and constraint and compliance checks.

8. The adaptive autonomous driving decision-making method integrating user preferences and traffic rules according to claim 1, characterized in that, In step 5, the weights in the comprehensive cost function are adjusted based on the multi-objective weight adjustment vector matched in step 3 or generated in step 4, specifically as follows: Based on the first in the comprehensive cost function d The current weight vector corresponding to each target With multi-objective weight adjustment vector The weights of each objective are normalized and updated to ensure that the sum of the weights of all objectives remains 1, reflecting their relative importance. The update formula is as follows: Wherein, the multi-objective weight adjustment vector The multi-objective weight adjustment vector matched in step 3 Or the multi-objective weight adjustment vector generated in step 4 The first in d Each weight adjustment value, For the first d The adjusted weight values ​​for each objective.

9. The adaptive autonomous driving decision-making method integrating user preferences and traffic rules according to claim 1, characterized in that, In step 5, generating the optimal path based on the adjusted comprehensive cost function specifically involves: A sampling-based path planning method is used to calculate the comprehensive cost function value of each candidate path. This comprehensive cost function considers five key objectives: safety, efficiency, longitudinal comfort, lateral comfort, and path deviation. + + in, Indicates the first The comprehensive cost function value of the candidate paths; These are the adjusted normalized weights corresponding to each objective function; For the first The safety cost function of a path is defined as the collision risk with surrounding traffic participants, which is quantified and assessed using Mahalanobis distance. The path is Time and surrounding traffic participants Collision probability Defined as: in, Indicates the relationship between the vehicle and surrounding road users. At any moment The square of the Mahalanobis distance, For numerical stability; No. The safety cost function of a path is expressed as follows during the planning period. Within, for all observable surrounding traffic participants The result of summing and normalizing the collision probabilities: in, Represents the normalization function; For the first The efficiency-cost function of a path is expressed as the average speed versus the speed limit. The normalized result of the bias is defined as: in, Indicates the first The path at time speed; , The first The lateral and longitudinal comfort cost functions for a path, represented by the acceleration change rate jerk (Jerk), are defined as the normalized result of the cumulative Jerk values ​​in the lateral and longitudinal directions: in, , They represent the first The path at time The horizontal and vertical Jerk values; For the first The path deviation cost function for each path is expressed as the average weighted lateral deviation between the vehicle's path and the reference path: in, A set representing the lateral offset of a path. For the first The lateral offset of the path endpoint. This indicates the number of sampling points contained in the path, i.e., the path length; Multiple candidate path samples are generated within each planning cycle, based on the updated weight parameters. Calculate their comprehensive cost functions respectively. The path with the lowest cost is selected as the optimal path.

Images