Automatic driving layered autonomous decision planning method and system based on lane gap attention characteristics

Abstract

The invention discloses an automatic driving layered autonomous decision planning method and system based on lane gap attention characteristics, and the method comprises the steps: collecting BEV images in a vehicle manual driving simulation scene and corresponding lane gap images, and constructing a training set; and lane gap segmentation and attention feature extraction are carried out based on aerial view features, and a CILQR optimizer is adopted to update a vehicle reference line, adjust a target speed and carry out vehicle motion planning based on an action decision and a vehicle motion process of each unmanned vehicle. The understanding of the scene is improved by learning the attention of the lane gap, and safe and efficient autonomous decision making and planning are further achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent vehicles, and in particular relates to a method and system for hierarchical autonomous decision-making planning of automatic driving based on lane gap attention characteristics. Background Art

[0002] In the past decade, autonomous vehicle technology has made significant progress and has become one of the hot spots in global technology research and development. As urban traffic environments become increasingly complex, traditional rule-based autonomous driving decision-making methods have gradually exposed their limitations and are unable to cope with various dynamic and uncertain driving scenarios, including:

[0003] Lack of robustness: Traditional rule-based or model-based decision-making methods often lack robustness when faced with complex and dynamically changing traffic environments. For example, these methods may not be able to make adaptive adjustments when dealing with unexpected situations or non-standard traffic behaviors.

[0004] Limited generalization ability: Many existing methods have limited generalization ability across different scenarios. They may perform well in the training environment, but the performance may drop significantly in unseen scenarios.

[0005] Insufficient scene understanding: Many methods lack a deep understanding of the driving scene, especially in identifying and predicting the intentions of other traffic participants. This may lead to unsafe or inefficient decisions in complex traffic scenarios.

[0006] Reinforcement learning, an artificial intelligence method that can learn the best strategy through trial and error, has begun to show great potential in the field of autonomous driving. Especially in complex urban scenarios, vehicles need to respond flexibly in different lane gaps to achieve safe and efficient driving. Traditional methods often have difficulty maintaining efficiency under changing traffic conditions, while decision-making methods based on reinforcement learning can perform well in diverse driving environments through continuous learning and optimization. However, although reinforcement learning has shown its application potential through trial and error learning, its generalization is limited due to insufficient scene understanding.

[0007] Therefore, it is particularly important to study a more intelligent and flexible decision-making planning method. Summary of the invention

[0008] To solve the above problems, the present invention provides a method and system for hierarchical autonomous decision-making and planning for autonomous driving based on lane gap attention features, which improves the understanding of the scene and further realizes safe and efficient autonomous decision-making and planning.

[0009] The technical solution to achieve the purpose of the present invention is:

[0010] A hierarchical autonomous decision-making planning method for autonomous driving based on lane gap attention features, comprising:

[0011] Step 1, collecting BEV images and corresponding lane gap images in a vehicle manual driving simulation scenario to construct a training set;

[0012] Step 2, train the image segmentation network through the training set, obtain the lane gap feature through the trained image segmentation network, combine the lane gap feature with the BEV image to obtain the extended feature;

[0013] Step 3: Perform cross-attention operations on the extended features and the main vehicle status features through a cross-attention network to obtain multi-dimensional comprehensive features;

[0014] Step 4: Based on the multi-dimensional comprehensive features, the trained reinforcement learning algorithm network is used to obtain the action decision of each unmanned vehicle;

[0015] Step 5: Based on the action decisions and vehicle motion process of each unmanned vehicle, the CILQR optimizer is used to update the vehicle reference line, adjust the target speed, perform vehicle motion planning, and feed the optimization results back to the simulation environment to update the experience recycling pool of the reinforcement learning network.

[0016] Furthermore, the input of the image segmentation network is the normalized BEV image feature The dimension is 3×h×w. The lane gap extraction network includes a multi-layer convolution upsampling and downsampling structure, in which the upsampling and downsampling layers both adopt a two-layer convolution structure, with a total of three upsampling and downsampling. The output lane gap feature dimension is 1×h×w, where h and w represent the height and width of the image.

[0017] Furthermore, step 3 specifically includes: the cross attention network receives the main vehicle state feature As a query, expand the features As the key Key and value Value, the attention weight is generated by calculating the similarity between the query and the key, and the value Value is weighted summed using the calculated attention weight to generate the output Expand the feature With cross-attention feature Combination to obtain multi-dimensional comprehensive features Among them, the main vehicle state characteristics Including speed, acceleration, heading angle and wheel angle.

[0018] Furthermore, the multi-dimensional comprehensive features are extracted using a convolutional network, and the main vehicle state features are extracted using a multi-layer perceptron. The two features are concatenated and used as input to a reinforcement learning algorithm network. The action decisions output by the reinforcement learning algorithm network are discrete actions, including changing lanes to the left, keeping lanes, changing lanes to the right, accelerating and decelerating, which are input to the CILQR optimizer.

[0019] Furthermore, the reinforcement learning algorithm network is trained and updated by a mean square error loss function, and the mean square error loss function is updated by ∑R t Build, R t is the reward value corresponding to the autonomous driving vehicle.

[0020] Furthermore, the reward value R t for:

[0021]

[0022] Where, v represents the current speed of the vehicle, v max is the maximum speed of the vehicle, and τ is the speed reward ratio.

[0023] Furthermore, the reinforcement learning algorithm network represents the intention decision process with a tuple: M(S, A, P, r, γ), wherein the state space S is used to describe the current driving state and environmental state of the vehicle, the action space A represents the upper-level decision instructions of the vehicle, and is represented as the state of the decision target at the next moment; r represents the reward function, which rewards safe driving and task completion and punishes collisions; γ is a discount factor for future rewards.

[0024] Furthermore, the vehicle movement process is:

[0025]

[0026] stx k+1 =f(x k ,u k ),k=0,1,…,N-1

[0027] x 0 =x start

[0028]

[0029] u min k max ,k=1,2,…,N-1

[0030] Among them, φ(x N ) represents x N The cost function under the state, L k ​​represents the action cost function at time k, f(x k ,u k ) is the state update model, x k represents the vehicle state at time k, x 0 、x start is the initial state of the vehicle, u k represents the control input of the vehicle at time k, u min 、u max is the minimum and maximum value set; is the obstacle collision constraint, where Represents the state of the j-th obstacle at time k.

[0031] Furthermore, the action cost function includes vehicle control cost and vehicle state cost, wherein:

[0032] The vehicle control cost is:

[0033]

[0034] Among them, w acc 、w steer Represent the acceleration and corner cost parameters respectively;

[0035] The vehicle state cost includes a speed following cost Follow the reference line The speed follows the cost for:

[0036]

[0037] Among them, x k is the vehicle state at time k, v r is the reference speed, w vel is the speed following cost parameter;

[0038] The reference line following cost for:

[0039]

[0040] Among them, ref k Represented as trajectory point x k The projection point to the target lane reference line, dis(x k ,ref k ) is the distance function, w ref is the reference line following cost parameter.

[0041] An autonomous driving hierarchical decision-making planning system based on lane gap attention features, comprising:

[0042] The lane gap segmentation and attention feature extraction unit includes a data acquisition unit and an image feature segmentation unit. The data acquisition unit collects BEV images and corresponding lane gap images in a manual driving scenario of a vehicle to construct a training set. The image feature segmentation unit trains an image segmentation network through the training set, obtains lane gap features through the trained image segmentation network, combines the lane gap features with the BEV image to obtain extended features, and performs cross-attention operations on the extended features and the main vehicle state features through a cross-attention network to obtain multi-dimensional comprehensive features.

[0043] The upper-level behavior decision-making unit uses a trained reinforcement learning algorithm network to obtain the action decisions of each unmanned vehicle based on multi-dimensional comprehensive features;

[0044] The lower-level trajectory planning unit uses the CILQR optimizer to update the vehicle reference line, adjust the target speed, perform vehicle motion planning, and feed the optimized data back to the data acquisition unit for updating the reinforcement learning algorithm network.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] 1. The present invention provides a hierarchical autonomous decision-making planning method for autonomous driving based on lane gap attention features. The method includes three aspects: an autonomous driving multi-scenario feature extraction method based on BEV features, deep reinforcement learning is used to learn autonomous driving strategies, upper-level decisions are generated, and a hierarchical planning framework is used. The lower layer uses a CILQR optimizer to plan trajectories, thereby achieving generalization of complex scenarios.

[0047] 2. The present invention provides a multi-scenario feature extraction method for autonomous driving based on BEV features. The BEV image is used to segment the lane gaps of interest to the main vehicle and combine them into extended features. The cross-attention of the main vehicle state and the extended features is adopted, and the acquired attention features are combined with the extended features to obtain more comprehensive features, perform multi-layer convolution processing, and enhance the understanding of the driving scene.

[0048] 3. The present invention adopts the above-mentioned feature extraction network, cooperates with the PPO reinforcement learning framework, and designs an upper and lower layered framework to provide a unified solution to the autonomous driving planning problem, thereby realizing the network's adaptation to various scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is the multi-scenario BEV characteristic diagram of the present invention.

[0050] Figure 2 Schematic diagram of lane gap extraction according to the present invention.

[0051] Figure 3 It is the lane gap segmentation network design of the present invention.

[0052] Figure 4 This is a training block diagram of the system of the present invention.

[0053] Figure 5 Schematic diagram of the reference line for tracking the autonomous driving vehicle of the present invention.

[0054] Figure 6 This is a multi-scenario rendering of the autonomous driving of the present invention. DETAILED DESCRIPTION

[0055] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0056] The present invention provides an autonomous driving hierarchical decision-making planning system based on lane gap attention characteristics, which includes three parts:

[0057] Lane gap segmentation and attention feature extraction layer based on Bird's Eye View (BEV) features;

[0058] The upper behavior decision layer based on reinforcement learning;

[0059] The lower trajectory planning layer is based on the Constrained Iterative Linear Quadratic Regulator (CILQR) optimization method.

[0060] The extraction layer selects the driving intention or goal according to the current environmental state, the lower trajectory planning layer plans the trajectory, and the upper behavior decision layer completes the trajectory tracking, that is, completes the execution of an action, and the system transfers to the state at the next moment.

[0061] The process of image data processing in the lane gap segmentation and attention feature extraction layer includes:

[0062] S1-1: Use the simulation environment to collect BEV images in the manual driving scenario, represented as 3-channel RGB images, such as Figure 1 As shown in the figure; the BEV image includes lane information within the detection range; the main vehicle position; the obstacle vehicle's continuous 5-frame trajectory, which adjusts the transparency of the trajectory according to time; the main vehicle routing path, which consists of a set of dense navigation points at the lane level;

[0063] Collect the lane gap (Gap) image corresponding to the BEV image, where the lane gap (Gap ij ) is defined as the gap along the lane between vehicles (i, j) or between vehicle (i) and a static obstacle (j), such as Figure 2As shown in the figure; the samples required in training are to accurately extract the lane structure, and to perform targeted lane extraction and labeling based on the known road network structure and accurate road information; in the actual autonomous driving process, it is impossible to accurately extract lane gaps for different structured roads, so it is of great significance to use a deep learning-based solution to extract lane gaps.

[0064] S1-2: The present invention uses the collected BEV image and Gap image as training data and label data to perform image segmentation network (Net gap ) supervised learning training, such as Figure 3 shown.

[0065] S1-3: BEV image is normalized and used as lane gap segmentation network Net gap The lane gap feature is obtained by taking the input of . The BEV image feature is a 3-channel RGB image, which is normalized to be the BEV feature at time t. The dimension is 3×h×w, and the lane gap feature obtained The dimension is 1×h×w. The two features are concatenated to obtain a new extended feature. The dimensions are 4×h×w.

[0066] S1-4: The state of the main vehicle at time t is expressed as They represent speed, acceleration, heading angle and wheel angle respectively.

[0067] S1-5: For extended features Characteristics of the vehicle status Perform a cross-attention operation to associate and integrate information from two different sources. The cross-attention network receives the main vehicle state features. As a query, expand the features As the key and value, attention weights are generated by calculating the similarity between the query and the key. These weights reflect the relevance between the query and the key. The value is weighted summed using the calculated attention weights to generate the output Expand the feature With cross-attention feature Combination to obtain multi-dimensional comprehensive features

[0068] The second module of the present invention, the upper behavior decision layer, uses the reinforcement learning algorithm Proximal Policy Optimization (PPO) to train and optimize the driving strategy. The training process and steps are as follows:

[0069] S2-1: The intention decision process is represented by a tuple M(S, A, P, r, γ) using the PPO algorithm. The state space S is used to describe the current driving state and environmental state of the vehicle, which can be composed of the road structure in the environment and the motion state vector of the vehicle. The action space A represents the upper-level decision instructions of the vehicle, which are usually expressed as the state of the decision target at the next moment, including lateral and longitudinal positions and the expected speed. On urban roads, the lateral position can be roughly represented by the target lane, and the speed information can be represented by instructions such as overtaking, deceleration, and acceleration. r represents the reward function, which rewards safe driving and task completion and penalizes collisions; γ is the discount factor for future rewards.

[0070] S2-2: Initialize the training environment and use the BEV image and the main vehicle status of the training environment as the observation Obs at the current time t t , input into the feature extraction network designed in module 1, obtain the combined features and input into the PPO algorithm network to obtain the action decision of the unmanned vehicle. In the present invention, discrete actions are used as the output of the PPO strategy network, including {Left, Keep, Right, Speed_up, Speed_down}, which respectively represent changing lanes to the left, keeping lanes, changing lanes to the right, accelerating, and decelerating.

[0071] S2-3: The upper layer instructions are used as inputs to the lower layer CILQR planner to perform trajectory planning. The vehicle executes according to the planned trajectory and obtains a single-step reward R t , t represents the tth step. The single-step reward is expressed as:

[0072]

[0073] The vehicle is rewarded for reaching the target point and penalized for collision, where v represents the current speed of the vehicle and v max is the maximum speed of the vehicle, and τ is the speed reward ratio, which is set to 0.8.

[0074] S2-4: Based on the reward value R corresponding to each autonomous driving vehicle t Calculate the long-term reward by using the difference in value functions obtained by the critic network of the current strategy and the old strategy as the advantage function value, and calculate the probability ratio of taking each state action under the new strategy and the old strategy, that is, the ratio, to measure the improvement of the new strategy over the old strategy.

[0075] S2-5: Calculate the agent objective function and use the ratio and advantage function to construct the agent objective function, which is the expected return under the new strategy with pruning terms. The existence of pruning terms can limit the range of changes in the new strategy and ensure the stability of strategy updates.

[0076] S2-6: Use the new strategy and repeat the above steps until convergence or the specified number of training rounds is reached.

[0077] The present invention aims at the problem of multi-scenario autonomous driving in cities and divides the autonomous driving planning problem into upper-level decision-making and lower-level trajectory planning modules. After receiving the decision action output by the PPO strategy, the CILQR optimizer updates the vehicle reference line, adjusts the target speed, and performs planning. The third module of the present invention, the lower-level trajectory planning layer, specifically includes:

[0078] S3-1: The motion process of the vehicle is expressed as:

[0079]

[0080] stx k+1 =f(x k ,u k ),k=0,1,…,N-1

[0081] x 0 =x start

[0082]

[0083] u min k max ,k=1,2,…,N-1

[0084] Among them, φ(*) represents the cost function in the * state, L k represents the action cost function at time k. f(x k ,u k ) is the state update model, x k represents the vehicle state at time k, u k represents the control input of the vehicle at time k. In the planning process, it is necessary not only to plan a driving trajectory that meets the driving intention according to the upper-level decision instructions, but also to meet the vehicle constraints, including the vehicle controller constraints u min k max , obstacle collision constraints in, Represents the state of the j-th obstacle at time k.

[0085] S3-2: In the CILQR algorithm, the planning objective function L k (x k ,u k ) including vehicle control costs and vehicle state cost The control cost is divided into acceleration cost and steering angle cost, and the state cost is divided into speed following cost.​​​​ Follow the reference line And design reasonable vehicle constraints, including acceleration constraints, steering angle constraints and obstacle avoidance constraints.

[0086] The PPO strategy network outputs discrete action values, where the lane change decision is used by the CILQR optimizer to update the reference line of the autonomous driving vehicle, and the speed decision is used to update the target speed of the autonomous driving vehicle; design the control cost function as follows:

[0087]

[0088] Among them, w acc 、w steer Respectively represent the acceleration and steering angle cost parameters, which are set to 0.1 and 1.0 in this embodiment. The acceleration cost is related to energy loss and comfort, and penalizes excessive acceleration and deceleration actions. The steering angle cost penalizes excessive vehicle steering. as follows:

[0089]

[0090] Among them, x k is the vehicle state at time k, v r is the reference speed, w vel is the speed following cost parameter, which is set to 1.0 in this embodiment. In the present invention, the target speed is given by the upper layer decision to guide the longitudinal travel of the autonomous driving vehicle. It is expressed as:

[0091]

[0092] Among them, ref k Represented as trajectory point x k The projection point to the target lane reference line, dis(*) is the distance function, w ref is the reference line following cost parameter, which is set to 6.0 in this embodiment. The purpose of introducing this cost is to guide the vehicle to approach the target reference line, so as to achieve the specified driving goal, such as Figure 5 shown.

[0093] S3-3: The obtained CILQR optimized trajectory is used as the state update of the simulation environment. The state of the autonomous driving vehicle is updated at a frequency of 10Hz to obtain the reward R t And the updated observations Obs t+1 , thereby updating the experience recycling pool (replay buffer) of the PPO reinforcement learning network of the second module.

[0094] Example 1

[0095] The present invention provides a method for hierarchical autonomous decision-making planning of autonomous driving based on lane gap attention characteristics, which is characterized by comprising:

[0096] Step 1: collect BEV images and corresponding lane gap images in the vehicle manual driving simulation scene to build a training set; Figure 1 As shown, a BEV image is collected, the BEV image feature is 60*60 meters, the resolution is 0.2m, and the image size is 300*300. The center of the autonomous driving vehicle is located at the lateral center position, the forward field of view is 40 meters, and the rearward field of view is 20 meters.

[0097] S2: Train the image segmentation network through the training set, obtain the lane gap feature through the trained image segmentation network, combine the lane gap feature and the BEV image to obtain the extended feature; including:

[0098] S2-1, randomly initialize the multi-scenario simulation environment, obtain the BEV image and the main vehicle motion state as the observation output. The BEV image includes the lane information within the detection range; the main vehicle position; the obstacle vehicle's continuous 5-frame trajectory, which adjusts the transparency of the trajectory according to time; the main vehicle routing information, that is, a set of dense navigation points at the lane level. The main vehicle state is represented as They represent speed, acceleration, heading angle and wheel angle respectively.

[0099] S2-2, BEV image feature is a 3-channel RGB image, which is the BEV feature at time i after normalization The dimensions are (3×h×w), through Figure 2 Pre-trained lane gap extraction network, obtained lane gap features The dimension is (1×h×w), and the two features are combined to obtain a new extended feature The dimensions are (4×h×w).

[0100] In the present invention, the lane gap extraction network includes a multi-layer convolution upsampling and downsampling structure, wherein both the upsampling and downsampling layers adopt a two-layer convolution structure, with a total of three upsampling and downsampling.

[0101] S3, cross-attention operations are performed on the extended features and the main vehicle status features through the cross-attention network to obtain multi-dimensional comprehensive features.

[0102] The state of the main vehicle at time i is expressed as Represent the speed, acceleration, heading angle and wheel turning angle respectively. Figure 4 As shown, for the extended features Characteristics of the vehicle status Perform cross-attention operations to associate and integrate information from two different sources. The cross-attention network receives the state features of the main vehicle. As a query, expand the features As keys and values, attention weights are generated by calculating the similarity between the query and the key. The values ​​are weighted summed using the calculated attention weights to generate the output Expand the feature With cross-attention feature Combination to obtain multi-dimensional comprehensive features

[0103] S4: The multi-dimensional comprehensive features are extracted using a convolutional network, and the main vehicle state is extracted using a multi-layer perceptron. The two features are concatenated and used as the input of the PPO strategy network to obtain the action decisions of each unmanned vehicle.

[0104] S5: Based on the action decisions and vehicle motion process of each unmanned vehicle, the CILQR optimizer is used to update the vehicle reference line, adjust the target speed, perform vehicle motion planning, and feed the optimization results back to the simulation environment to calculate the reward value R i , reward the vehicle for reaching the target point and punish the collision; and then update the experience recycling pool of the reinforcement learning network.

[0105] S6: Update the data storage space, set i=i+1, and repeat S1 to S5.

[0106] When training the network model, the following network update steps should also be performed:

[0107] S7: According to the reward value R corresponding to the autonomous driving vehicle i The sum of ∑R i Construct the mean square error loss function, and then update the PPO network according to the mean square error loss function;

[0108] S8: Use the updated PPO network to re-execute steps S1 to S7 until convergence or the specified number of training rounds is reached. Figure 6 As shown, in the unprotected left turn and merging lane scenarios, the autonomous driving vehicle independently decides the speed and lane change intention to achieve safe trajectory planning.

[0109] Of course, the present invention may have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art may certainly make various corresponding changes and modifications based on the present invention, but these corresponding changes and modifications should all fall within the scope of protection of the claims attached to the present invention.

Claims

1. A hierarchical autonomous decision-making planning method for autonomous driving based on lane gap attention features, characterized in that: include: Step 1, collecting BEV images and corresponding lane gap images in a vehicle manual driving simulation scenario to construct a training set; Step 2, train the image segmentation network through the training set, obtain the lane gap feature through the trained image segmentation network, combine the lane gap feature with the BEV image to obtain the extended feature; Step 3: Perform cross-attention operations on the extended features and the main vehicle status features through a cross-attention network to obtain multi-dimensional comprehensive features; Step 4: Based on the multi-dimensional comprehensive features, the trained reinforcement learning algorithm network is used to obtain the action decision of each unmanned vehicle; Step 5: Based on the action decisions and vehicle motion process of each unmanned vehicle, the CILQR optimizer is used to update the vehicle reference line, adjust the target speed, perform vehicle motion planning, and feed the optimization results back to the simulation environment to update the experience recycling pool of the reinforcement learning network.

2. The method for hierarchical autonomous decision-making and planning for autonomous driving based on lane gap attention features according to claim 1, characterized in that: The input of the image segmentation network is the normalized BEV image feature The dimension is 3×h×w. The lane gap extraction network includes a multi-layer convolution upsampling and downsampling structure, in which the upsampling and downsampling layers both adopt a two-layer convolution structure, with a total of three upsampling and downsampling. The output lane gap feature dimension is 1×h×w, where h and w represent the height and width of the image.

3. The method for hierarchical autonomous decision-making planning for autonomous driving based on lane gap attention features according to claim 1, characterized in that: The step 3 specifically includes: the cross attention network receives the main vehicle state feature As a query, expand the features As the key Key and value Value, the attention weight is generated by calculating the similarity between the query and the key, and the value Value is weighted summed using the calculated attention weight to generate the output Expand the feature With cross-attention feature Combination to obtain multi-dimensional comprehensive features Among them, the main vehicle state characteristics Including speed, acceleration, heading angle and wheel angle.

4. The method for hierarchical autonomous decision-making planning for autonomous driving based on lane gap attention features according to claim 1, characterized in that: The multi-dimensional comprehensive features are extracted using a convolutional network, and the main vehicle state features are extracted using a multi-layer perceptron. The two features are concatenated and used as the input of the reinforcement learning algorithm network. The action decisions output by the reinforcement learning algorithm network are discrete actions, including changing lanes to the left, keeping lanes, changing lanes to the right, accelerating and decelerating, which are the input of the CILQR optimizer.

5. The method for hierarchical autonomous decision-making planning for autonomous driving based on lane gap attention features according to claim 4, characterized in that: The reinforcement learning algorithm network is trained and updated by the mean square error loss function, and the mean square error loss function is updated by ∑R t Build, R t is the reward value corresponding to the autonomous driving vehicle.

6. The method for hierarchical autonomous decision-making planning for autonomous driving based on lane gap attention features according to claim 4, characterized in that: The reward value R t for: Where, v represents the current speed of the vehicle, v max is the maximum speed of the vehicle, and τ is the speed reward ratio.

7. The method for hierarchical autonomous decision-making planning for autonomous driving based on lane gap attention features according to claim 6, characterized in that: The reinforcement learning algorithm network represents the intention decision process with a tuple: M(S, A, P, r, γ), where the state space S is used to describe the current driving state and environmental state of the vehicle, the action space A represents the upper-level decision instructions of the vehicle, and is represented as the state of the decision target at the next moment; r represents the reward function, which rewards safe driving and task completion and punishes collisions; γ is a discount factor for future rewards.

8. The method for hierarchical autonomous decision-making planning for autonomous driving based on lane gap attention features according to claim 1, characterized in that: The vehicle motion process is: s.t.x k+1 =f(x k ,u k ),k=0,1,…,N-1 x0=x start u min <u k <u max ,k=1,2,…,N-1 Among them, φ(x N ) represents x N The cost function under the state, L k represents the action cost function at time k, f(x k ,u k ) is the state update model, x k represents the vehicle state at time k, x0, x start is the initial state of the vehicle, u k represents the control input of the vehicle at time k, u min 、u max is the minimum and maximum value set; is the obstacle collision constraint, where Represents the state of the j-th obstacle at time k.

9. The method for hierarchical autonomous decision-making planning for autonomous driving based on lane gap attention features according to claim 8, characterized in that: The action cost function includes vehicle control cost and vehicle state cost, where: The vehicle control cost is: Among them, w acc 、w steer Represent the acceleration and angle cost parameters respectively; The vehicle state cost includes a speed following cost Follow the reference line The speed follows the cost for: Among them, x k is the vehicle state at time k, v r is the reference speed, w vel is the speed following cost parameter; The reference line following cost for: Among them, ref k Represented as trajectory point x k The projection point to the target lane reference line, dis(x k ,ref k ) is the distance function, w ref is the reference line following cost parameter.

10. An autonomous driving hierarchical decision-making planning system for implementing any of the methods described in claims 1-9, characterized in that: include: The lane gap segmentation and attention feature extraction unit includes a data acquisition unit and an image feature segmentation unit. The data acquisition unit collects BEV images and corresponding lane gap images in a manual driving scenario of a vehicle to construct a training set. The image feature segmentation unit trains an image segmentation network through the training set, obtains lane gap features through the trained image segmentation network, combines the lane gap features with the BEV image to obtain extended features, and performs cross-attention operations on the extended features and the main vehicle state features through a cross-attention network to obtain multi-dimensional comprehensive features. The upper-level behavior decision-making unit uses a trained reinforcement learning algorithm network to obtain the action decisions of each unmanned vehicle based on multi-dimensional comprehensive features; The lower-level trajectory planning unit uses the CILQR optimizer to update the vehicle reference line, adjust the target speed, perform vehicle motion planning, and feed the optimized data back to the data acquisition unit for updating the reinforcement learning algorithm network.

Images