CN-121977601-A - Decision planning method based on neural network heuristic mixed A-search and parallel optimization

CN121977601ACN 121977601 ACN121977601 ACN 121977601ACN-121977601-A

Abstract

The invention provides a decision planning method based on neural network heuristic hybrid A search and parallel optimization, which relates to the technical field of automatic driving decision planning, and aims to acquire and preprocess real-time environment perception data, generate heuristic information based on the neural network, initialize a hybrid A search algorithm, optimize an evaluation function, embed the neural heuristic network in the hybrid A search, construct a node feature vector, combine vehicle kinematic constraint and collision risk verification to generate a feasible path, realize parallel search and optimization through multithreading or GPU acceleration, screen an optimal path according to smoothness, safety, comfort and time efficiency, generate a control instruction according to the optimal path, monitor a tracking effect in real time and feed back to the network, dynamically adjust a model and a strategy, realize continuous optimization, combine parallel optimization through the neural network and the heuristic search, improve planning speed and quality, and enhance adaptability, instantaneity and decision accuracy in a complex environment.

Inventors

HU JUN
QU XIN
YAO YITONG
CHEN TIANTIAN

Assignees

沈阳工业大学

Dates

Publication Date: 20260505
Application Date: 20260210

Claims (5)

1. A decision planning method based on neural network heuristic mixed A-search and parallel optimization is characterized by comprising the following steps: Step S1, acquiring real-time environment perception data, preprocessing the real-time environment perception data, and constructing a fusion environment representation containing static road information and dynamic obstacle prediction information; Step S2, learning environmental features and driving modes from historical driving data by utilizing a neural network to generate heuristic information, initializing a mixed A search algorithm, defining a starting point and a target point, dividing a search space, and optimizing an evaluation function by using the heuristic information to obtain an optimized mixed A search algorithm; Step 3, based on the fusion environment representation, embedding a trained neural heuristic network in the optimized mixed A search, constructing a multidimensional feature vector of the node to be expanded, obtaining the comprehensive evaluation cost of the node, combining the vehicle kinematic constraint and collision risk verification, pruning and optimizing the search space, and generating one or more feasible paths; Step S4, parallel processing of path searching and optimizing is realized through a multithreading technology or GPU acceleration, and task division is optimized through a heuristic algorithm; And step S5, generating steering, accelerating and braking control instructions according to the screened optimal path, monitoring path tracking errors, vehicle stability and safety indexes in real time, using an execution result as feedback information to update an evaluation input or experience sample library of the neural network, and adaptively adjusting heuristic search strategies and cost weights to realize continuous optimization on the premise of not changing network structure parameters.
2. The decision-making planning method based on neural network heuristic hybrid a-search and parallel optimization of claim 1, wherein the specific method for acquiring and preprocessing real-time environmental awareness data in step S1 to construct a fusion environmental representation including static road information and dynamic obstacle prediction information comprises the following steps: the method for acquiring the real-time environment sensing data comprises the following steps: A camera, a laser radar and a millimeter wave radar sensor equipped on the vehicle, which captures detailed data of the surrounding environment in real time; the camera is responsible for capturing high-resolution visual information and is used for identifying traffic signs, signal lamps and pedestrians; The laser radar creates a 3D environment map by emitting laser pulses and measuring reflection time for detecting obstacles and road boundaries; the millimeter wave radar can penetrate severe weather conditions such as fog, rain and the like by utilizing electromagnetic waves with longer wavelength, and provides long-distance obstacle detection; preprocessing acquired real-time environment perception data, and constructing a fusion environment representation comprising static road information and dynamic obstacle prediction information, wherein the method comprises the following steps: Eliminating noise of the environment sensing data through a filtering technology; integrating information from different sensors by utilizing a data fusion technology to form a consistent environment representation; the preprocessing also comprises the integration of static map information, so that the current geographic position and road layout are known, and meanwhile, the dynamic track of the obstacle is predicted so as to plan a path.
3. The decision-making planning method based on neural network heuristic hybrid A-search and parallel optimization of claim 1, wherein in step S2, the method for generating heuristic information by learning environmental features and driving modes from historical driving data by using the neural network comprises the following steps of: Based on historical driving data, utilizing neural network to learn environmental characteristics and driving mode to make training so as to obtain a model for node forward search, after the neural network training based on the historical driving data is completed, using said model as cost/feasibility evaluation module of mixed A forward expansion, firstly, making algorithm initialization, setting vehicle kinematic model and search parameters, and setting the search parameters including step length, steering discrete number, maximum curvature/minimum turning radius, reversing switch and collision detection resolution, and constructing priority queue Open List, closed Set and cost mapping, then defining complete state quantity and position of starting point and target point under the map coordinate system according to the task Heading and heading direction Speed of The continuous space is rasterized according to the position and the course is discretized into grids, and meanwhile, a search boundary is determined in a local planning range and feasible motion primitives are generated for forward expansion of the nodes; The evaluation function adopts , wherein, Accumulating driving distance, steering/shifting, curvature change and obstacle attaching, Heuristic information is introduced, convergence speed and path quality are improved through weighted combination and consistency constraint, and therefore the number of expansion nodes is obviously reduced while feasibility is guaranteed.
4. The decision-making planning method based on neural network heuristic mixed a-search and parallel optimization according to claim 1, wherein in step S3, based on the fusion environment representation, a trained neural heuristic network is embedded in the optimized mixed a-search, a multidimensional feature vector of a node to be expanded is constructed, a node comprehensive evaluation cost is obtained, and a vehicle kinematic constraint and collision risk verification are combined, and a specific method for pruning and optimizing a search space to generate one or more feasible paths comprises: In the path searching process, a mixed A-type searching algorithm is adopted, and a neural heuristic network is introduced in a node expansion stage to guide the searching process; for each node ni to be expanded in the search space, constructing a corresponding node feature vector according to the current motion state parameters of the vehicle, road structure information and a prediction result of the dynamic obstacle: ; Wherein, (x i ,y i ,θ i ) represents a node pose, v i represents a vehicle speed, a i represents a vehicle acceleration, k i represents a path curvature, r i represents a road structural feature, and o i represents a dynamic obstacle prediction feature; inputting the node characteristic vector X i into a pre-trained neural heuristic network The remaining cost estimation value h i from the node to the target state is output, namely: h i =H(X i ) Taking the residual cost estimated value h i as a heuristic cost item, and forming the comprehensive estimated cost f i of the node together with the accumulated path cost g i of the node, wherein the accumulated path cost g i of f i =g i +λh i is formed by weighting a plurality of child cost items: ; ; ; ; ; Where DeltaS k denotes the path length between adjacent nodes, K k denotes the curvature, o j (t) denotes the dynamic obstacle position at the predicted time t, A distance penalty function; Sorting and selecting candidate nodes based on the comprehensive evaluation cost f i to determine an expansion sequence of the nodes, so as to guide the mixed A-type search process to a search direction with higher feasibility and safety; in the node expansion process, carrying out feasibility verification on the generated candidate nodes by combining with vehicle kinematic constraint conditions, wherein the vehicle kinematic model meets the following conditions: ; meanwhile, the following constraint conditions are applied to the candidate nodes: ; and removing the node d (n i ,o j (t))<d safe ) with collision risk according to the dynamic obstacle prediction result, gradually completing pruning and optimizing the search space, and finally generating a high-efficiency feasible path which meets the dynamic obstacle constraint and the vehicle state constraint and is suitable for the actual driving scene.
5. The decision planning method based on neural network heuristic hybrid a-search and parallel optimization of claim 1, wherein in step S4, the parallel processing of path search and optimization is realized by a multithreading technology or GPU acceleration, task division is optimized by using a heuristic algorithm, and the specific method for screening the optimal path from candidate feasible paths based on standards of smoothness, safety, comfort, time efficiency and the like comprises the following steps: in the parallelization path searching stage, a plurality of feasible path sets meeting vehicle kinematics and collision constraint are generated simultaneously through multithreading or GPU acceleration For selecting a solution with optimal comprehensive performance from candidate paths, introducing a path optimization cost function weighted by multiple indexes, and uniformly evaluating and sequencing each path; (1) For any candidate path The comprehensive cost is defined as: ; Wherein, the For the weight coefficient of each evaluation index, satisfy ; (2) Path synthesis cost function: ; Wherein, the For the curvature of the path it is, L represents the total length of the path; (3) Security cost: ; Wherein, the At the moment of the route And obstacles Is used for the distance of the minimum of (a), T represents the total time of planning a path; (4) Comfort costs; ; Wherein, the For the acceleration of the vehicle, Acceleration (jerk); the weight of the acceleration is determined by the weight of the acceleration, Representing jerk weights; (5) Time efficiency cost: ; Wherein, the For the total travel time of the path, For the path length of the optical fiber, The time-weighting coefficients are represented by a number of time-weighted coefficients, Representing a length weight coefficient; (6) Optimal path selection cost: 。

Description

Decision planning method based on neural network heuristic mixed A-search and parallel optimization Technical Field The invention relates to the technical field of automatic driving decision planning, in particular to a decision planning method based on neural network heuristic hybrid A-search and parallel optimization. Background In the field of automatic driving, a path planning technology is a key for ensuring safe and efficient running of a vehicle. Conventional path planning algorithms, such as Dijkstra and a, have been widely used in static environments. However, as autopilot technology evolves, the environment in which the vehicle is located becomes increasingly complex and dynamic, and conventional path planning algorithms tend to be frustrating in dealing with these environments. Particularly in a scene with strict real-time requirements, the traditional algorithm has obvious defects in the aspects of path calculation efficiency and response speed. To overcome these challenges, researchers have begun to explore hybrid approaches that combine neural networks with traditional path planning algorithms. Deep learning techniques, particularly in the field of image processing and natural language processing, have demonstrated their powerful ability to extract and process information in complex environments. The technology can train a model through a large amount of data, so that the model has the capability of automatically learning environmental characteristics and changing modes, and more intelligent and flexible support is provided for path planning. The existing path planning method gradually starts to adopt a neural heuristic mixed a search strategy. The method combines the self-adaptive learning capability of the neural network and the path searching efficiency of the A algorithm, so that the path planning can not only quickly respond to the change in the dynamic and complex environment, but also generate a path solution with better quality and more adaptability. In order to further improve the calculation efficiency, parallel optimization technology is introduced, so that the algorithm can process a large amount of data and multiple path selection schemes in a short time, and the real-time response capability of the system is ensured. However, the existing methods still have some limitations, especially how to effectively integrate heuristic information generated by the neural network and a path selection strategy of the traditional search algorithm in the path planning process. The current model is likely to face the problems of untimely updating of information, sub-optimal path selection and the like when processing large-scale and rapid-change environment data. Therefore, further research on how to optimize the structural design of the neural network heuristic hybrid path planning model improves the response speed and the decision accuracy of the neural network heuristic hybrid path planning model in a dynamic environment, and becomes an important subject to be solved in the field of automatic driving. Disclosure of Invention The invention provides a decision planning method based on neural network heuristic mixed A-search and parallel optimization, aiming at solving the problems of high computational complexity, slow response speed and poor adaptability in a complex environment in the traditional decision planning method. The invention provides a decision planning method based on neural network heuristic mixed A search and parallel optimization, which comprises the following steps: Step S1, acquiring real-time environment perception data, preprocessing the real-time environment perception data, and constructing a fusion environment representation containing static road information and dynamic obstacle prediction information; Step S2, learning environmental features and driving modes from historical driving data by utilizing a neural network to generate heuristic information, initializing a mixed A search algorithm, defining a starting point and a target point, dividing a search space, and optimizing an evaluation function by using the heuristic information to obtain an optimized mixed A search algorithm; Step 3, based on the fusion environment representation, embedding a trained neural heuristic network in the optimized mixed A search, constructing a multidimensional feature vector of the node to be expanded, obtaining the comprehensive evaluation cost of the node, combining the vehicle kinematic constraint and collision risk verification, pruning and optimizing the search space, and generating one or more feasible paths; Step S4, parallel processing of path searching and optimizing is realized through a multithreading technology or GPU acceleration, and task division is optimized through a heuristic algorithm; And S5, generating steering, accelerating and braking control instructions according to the screened optimal path, monitoring path tracking effect in real time, feeding back an executio