Method and system for establishing and planning intelligent path of car in scenic spot based on AI algorithm

Abstract

The invention provides an AI algorithm-based scenic spot car intelligent path establishing and planning method and system, and relates to the technical field of path planning, and the method comprises the steps: collecting scenic spot road geographic information data, and constructing a weighted road network model; collecting real-time positioning of the trolley and sensor data; using the reinforcement learning model to generate a real-time optimal path planning strategy; calculating an initial recommendation path; road condition changes are monitored, and a path is dynamically adjusted; and outputting the updated optimal path scheme and carrying out navigation. According to the method, intelligent and real-time path planning of the vehicles in the scenic area can be realized.

Description

Technical Field

[0001] The present invention relates to path planning technology, and in particular to an AI algorithm-based intelligent path planning method and system for scenic area vehicles. Background Art

[0002] With the rapid development of tourism, traffic management in scenic spots is becoming increasingly complex. Traditional scenic spot car path planning methods mainly rely on static maps and fixed routes, which are difficult to adapt to the complex and changeable traffic conditions in scenic spots. Although GPS navigation technology has been widely used in recent years, it still has many shortcomings in the special environment of scenic spots.

[0003] First, existing path planning systems often ignore the particularity of scenic area roads. Scenic area roads are very different from urban roads. For example, factors such as road width, slope, and curvature have a more significant impact on the traffic capacity of small vehicles. However, current planning methods fail to fully consider these characteristics, resulting in the generated paths that may not be suitable for small vehicles in scenic areas.

[0004] Secondly, traditional path planning methods lack real-time and dynamic adjustment capabilities. Traffic conditions in scenic spots may change rapidly due to factors such as tourist flow and weather changes. Static path planning cannot reflect these changes in a timely manner, which may cause cars to get stuck in congestion or encounter temporarily closed roads, affecting the tourist experience.

[0005] Finally, existing technologies have not yet fully utilized artificial intelligence and big data technologies to optimize route planning. Although some systems have begun to use simple data analysis, they lack the support of advanced algorithms such as deep learning and reinforcement learning, and are unable to learn the optimal decision-making strategy from massive historical data and real-time information, making it difficult to achieve truly intelligent route planning.

[0006] These shortcomings limit the efficiency and service quality of scenic area vehicle traffic management, and a more intelligent, dynamic and adaptable path planning method is urgently needed to solve these problems. Summary of the invention

[0007] The embodiments of the present invention provide a method and system for establishing and planning an intelligent path for a scenic area vehicle based on an AI algorithm, which can solve the problems in the prior art.

[0008] According to a first aspect of the embodiments of the present invention,

[0009] Provides an AI-based algorithm-based intelligent path planning method for scenic area vehicles, including:

[0010] Collecting geographic information data of all roads in the scenic area, converting the geographic information data into a digital road network topology structure; performing feature extraction on the digital road network topology structure based on a deep learning neural network, extracting road width parameters, road slope parameters, road curvature parameters and road condition parameters; inputting the road width parameters, road slope parameters, road curvature parameters and road condition parameters into a preset road network evaluation model, generating a scenic area road capacity score, and establishing a weighted road network model including the road capacity score;

[0011] Collect real-time GPS positioning data and on-board sensor data of all cars in the scenic area, and obtain the current position information, driving speed information and driving direction information of the car based on the GPS positioning data; use the on-board sensor data to obtain obstacle distribution information and road congestion information around the car; input the current position information, driving speed information, driving direction information, obstacle distribution information and road congestion information into a preset reinforcement learning model, and combine with the weighted road network model to generate a real-time optimal path planning strategy;

[0012] Based on the real-time optimal path planning strategy, an initial recommended path is calculated; changes in road conditions on the initial recommended path are monitored in real time, and when changes in road conditions are detected, the changed road condition data is re-input into the reinforcement learning model; the reinforcement learning model evaluates and dynamically adjusts the current path in real time based on a path reward and punishment mechanism, outputs an updated optimal path plan, and sends the updated optimal path plan to the navigation system of the scenic area car for path navigation.

[0013] The method further comprises:

[0014] Road width information is collected through lidar point cloud data, and the RANSAC algorithm is used to extract the road surface boundary and calculate the width parameters. Sampling points are set every five meters on the road, and the sliding window method is used to perform local averaging on the sampling points to obtain a road width parameter matrix. The cubic spline interpolation method is used to construct a continuous slope curve based on elevation data, and equally spaced sampling is performed on the continuous slope curve to obtain discrete slope values ​​to form a road slope parameter matrix. The center line of the road is used to fit the arc through the least squares method to calculate the curvature radius, and the road curvature parameter matrix is ​​obtained.

[0015] Inputting the road width parameter, road slope parameter, road curvature parameter and road condition parameter into a preset road network evaluation model to generate a scenic area road capacity score, and establishing a weighted road network model including the road capacity score includes:

[0016] The ground image recognition results are combined with the vehicle-mounted sensor data to construct a road condition assessment index matrix including road surface smoothness parameters, friction coefficient parameters and damage degree parameters;

[0017] A width evaluation subnetwork is established for the road width parameter matrix, a slope evaluation subnetwork is established for the road slope parameter matrix, a curvature evaluation subnetwork is established for the road curvature parameter matrix, and a road condition evaluation subnetwork is established for the road condition evaluation index matrix; the evaluation subnetworks all use a soft attention mechanism to dynamically adjust feature weights, and achieve synergy of multiple key features through a multi-head attention mechanism;

[0018] The feature weights output by the evaluation subnetwork are weighted and summed with the corresponding standardized feature values ​​and environmental correction coefficients to obtain a road capacity score, wherein the environmental correction coefficient is updated in real time according to weather conditions and time period information through a Bayesian optimization method; the road capacity score adopts a standardized scale of 0-100;

[0019] A weighted road network graph structure is constructed based on the road capacity score, the road network node information is constructed as a vertex set, the road section information is constructed as an edge set, and the weight information based on the capacity score is constructed as an edge weight matrix; the edge weight matrix adopts a time-varying weight mechanism to dynamically adjust the basic score through time factors and weather factors; an exponential decay function is introduced in the edge weight calculation to simulate the cumulative effect of long-distance travel.

[0020] Inputting the current location information, driving speed information, driving direction information, obstacle distribution information and road congestion information into a preset reinforcement learning model, and combining with the weighted road network model to generate a real-time optimal path planning strategy includes:

[0021] The current position information of the vehicle is obtained through a positioning sensor, the driving speed information of the vehicle is obtained through a vehicle-mounted speed sensor, the driving direction information of the vehicle is obtained through a vehicle-mounted direction sensor, the obstacle distribution information is obtained through a vehicle-mounted laser radar scan, and the road congestion degree information is obtained through a road network monitoring device; the current position information, the driving speed information, and the driving direction information are mapped to a state space, and the obstacle distribution information and the road congestion degree information are mapped to an environment space;

[0022] Construct a dual Q learning network structure, including an evaluation network and a target network. The evaluation network inputs state space information and environment space information and outputs action value evaluation results. The target network calculates the target Q value based on historical evaluation data. Use a priority experience replay mechanism to store and sample training data, determine the priority of training samples according to the time difference error, and give priority to samples with larger time difference error for network update.

[0023] The action value evaluation result is integrated with the pre-built weighted road network model, and the path decision is sampled and evaluated based on the Monte Carlo tree search method. During the search process, the search tree is pruned and optimized in combination with the target Q value; a comprehensive score is calculated for each candidate path, and the comprehensive score includes three evaluation dimensions: path travel time, path safety, and path stability;

[0024] An optimal decision path is selected based on the comprehensive score, and a progressive dynamic programming method is used to locally optimize the optimal decision path. During the optimization process, vehicle steering constraints and acceleration and deceleration constraints are considered to ensure that the planned path meets the vehicle dynamics characteristics. The progressive dynamic programming method optimizes the path in segments through a rolling time domain to achieve real-time response to dynamic changes in the environment.

[0025] The optimal decision path is selected based on the comprehensive score, and the optimal decision path is locally optimized by using a progressive dynamic programming method. During the optimization process, the vehicle steering constraint and the acceleration and deceleration constraint are considered to ensure that the planned path meets the vehicle dynamic characteristics. The progressive dynamic programming method optimizes the path in segments through a rolling time domain to achieve real-time response to dynamic changes in the environment, including:

[0026] Establish a vehicle dynamics model, including vehicle center of mass motion state parameters, steering system parameters and power system parameters. Collect vehicle lateral acceleration, longitudinal acceleration, sideslip angular velocity, steering angular velocity and wheel speed data in real time through on-board sensors, and perform online identification and update of dynamics model parameters based on the extended Kalman filter algorithm.

[0027] Construct a path optimization objective function, and set path length minimization, steering energy minimization, and speed fluctuation minimization as optimization sub-objectives. At the same time, introduce vehicle steering angle constraints, steering angular velocity constraints, longitudinal acceleration constraints, and lateral acceleration constraints as hard constraints, and set the safe distance between the road boundary and the obstacle as a soft constraint.

[0028] The global planning path is segmented, and the optimization interval is defined by the sliding time window method. The optimization interval is divided into multiple equal time steps. The path curve is parameterized by a quintic polynomial in each time step to construct a state variable sequence including position, velocity, and acceleration.

[0029] For each optimization interval, a state transfer network is constructed based on the dynamic programming algorithm. The network nodes represent the feasible states of the vehicle, and the network edges represent the state transfer costs. During the state transfer process, the feasibility of the state transfer is verified through the vehicle dynamics model, and the state transfer edges that do not meet the constraints are eliminated.

[0030] A hierarchical search strategy is used to find the optimal path in the state transition network. First, a coarse-grained search is performed to determine the approximate search direction, and then a fine-grained search is performed near the optimal path to optimize the local trajectory. A heuristic evaluation function is introduced during the search process, combining the attraction of the target point with the repulsion of the obstacle as the evaluation basis.

[0031] The optimization interval is rolled and updated based on the progressive mechanism. The previous part of the current optimization interval is retained as the execution trajectory. The optimization window is slid forward by a fixed step size, and the state transfer network construction and path search process are repeated. During the interval update process, the path characteristics of adjacent optimization intervals are ensured to be continuous.

[0032] The reinforcement learning model evaluates and dynamically adjusts the current path in real time based on the path reward and punishment mechanism, outputs an updated optimal path plan, and sends the updated optimal path plan to the navigation system of the scenic spot car for path navigation, including:

[0033] Construct a multi-level path reward and punishment evaluation framework, divide the path evaluation indicators into safety indicators, efficiency indicators and comfort indicators, and calculate the scores of various evaluation indicators based on the driving data collected in real time by the on-board sensors; the safety indicators include the minimum distance to obstacles, the minimum distance to other vehicles, and the road curvature adaptability; the efficiency indicators include the average driving speed, the path length ratio, and the degree of congestion and detour; the comfort indicators include the acceleration fluctuation, the steering angle change rate, and the road surface smoothness;

[0034] A hierarchical reinforcement learning architecture is used to dynamically evaluate the path. The top-level policy network determines the path adjustment direction based on the global path characteristics, and the bottom-level execution network generates specific adjustment strategies based on local road condition information. In the evaluation process, spatiotemporal attenuation factors are integrated, and differentiated weights are given to historical evaluation data to highlight the impact of recent road condition changes.

[0035] Construct an adaptive reward and penalty function, where the reward items include the reduction in path travel time, the improvement in safety margin, and the improvement in comfort, and the penalty items include the degree of deviation from the original path, the loss of crossing congested areas, and the loss of frequent lane changes; the reward and penalty function adopts a dynamic weight mechanism to adaptively adjust the weight coefficients of each score according to the characteristics of the driving scene;

[0036] Based on the evaluation results of the reward and punishment function, the deep deterministic policy gradient algorithm is used to optimize the path plan, the path adjustment strategy is generated through the Actor network, and the strategy value is evaluated through the Critic network; the path smoothing constraint is introduced in the strategy generation process to ensure the continuity and feasibility of the path adjustment;

[0037] The optimized path plan is converted into a navigation instruction sequence, including the expected heading angle, expected speed, turning prompts and distance prompts, and the navigation instruction sequence is sent to the navigation system of the scenic area car through the on-board communication module; the navigation system performs real-time navigation based on the received navigation instruction sequence, and transmits the feedback information during the navigation process back to the path optimization module, forming a closed-loop optimization mechanism.

[0038] According to a second aspect of the embodiments of the present invention,

[0039] Provides an AI-based intelligent path planning system for scenic area vehicles, including:

[0040] The first unit is used to collect geographic information data of all roads in the scenic area, convert the geographic information data into a digital road network topology structure; perform feature extraction on the digital road network topology structure based on a deep learning neural network, extract road width parameters, road slope parameters, road curvature parameters and road condition parameters; input the road width parameters, road slope parameters, road curvature parameters and road condition parameters into a preset road network evaluation model, generate a scenic area road capacity score, and establish a weighted road network model including the road capacity score;

[0041] The second unit is used to collect real-time GPS positioning data and vehicle-mounted sensor data of all vehicles in the scenic area, obtain the current position information, driving speed information and driving direction information of the vehicle based on the GPS positioning data; use the vehicle-mounted sensor data to obtain obstacle distribution information and road congestion information around the vehicle; input the current position information, driving speed information, driving direction information, obstacle distribution information and road congestion information into a preset reinforcement learning model, and combine with the weighted road network model to generate a real-time optimal path planning strategy;

[0042] The third unit is used to calculate the initial recommended path based on the real-time optimal path planning strategy; monitor the changes in road conditions on the initial recommended path in real time, and when a change in road conditions is detected, re-input the changed road condition data into the reinforcement learning model; the reinforcement learning model evaluates and dynamically adjusts the current path in real time based on the path reward and punishment mechanism, outputs an updated optimal path plan, and sends the updated optimal path plan to the navigation system of the scenic area car for path navigation.

[0043] According to a third aspect of the embodiments of the present invention,

[0044] An electronic device is provided, comprising:

[0045] processor;

[0046] a memory for storing processor-executable instructions;

[0047] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0048] A fourth aspect of the embodiments of the present invention is:

[0049] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0050] The beneficial effects of this application are as follows:

[0051] 1. Improve the accuracy and adaptability of path planning:

[0052] The present invention collects and digitizes the geographic information of all roads in the scenic area, extracts key road parameters in combination with deep learning neural networks, and establishes a weighted road network model that includes road capacity scoring. This method can more comprehensively and accurately reflect the actual conditions of scenic area roads, and provide more reliable basic data for subsequent path planning, thereby improving the accuracy of path planning and its adaptability to actual road conditions.

[0053] 2. Realize real-time dynamic optimal path planning:

[0054] The present invention adopts a reinforcement learning model, combined with real-time collected GPS positioning data and on-board sensor data of the car, and can dynamically generate the optimal path planning strategy based on real-time information such as current traffic conditions, obstacle distribution, and road congestion. This method enables path planning to respond to changes in road conditions in a timely manner, providing the optimal driving route for scenic area cars, and improving driving efficiency and safety.

[0055] 3. Enhance the intelligence and adaptability of path planning:

[0056] The present invention monitors the changes in road conditions on the recommended path in real time, re-enters the changed data into the reinforcement learning model, and performs real-time evaluation and dynamic adjustment of the current path based on the path reward and punishment mechanism. This adaptive mechanism enables the path planning system to continuously learn and optimize, and the intelligence and accuracy of the system will continue to improve as the usage time increases. At the same time, by sending the updated optimal path plan to the navigation system of the scenic area car in real time, the timeliness and accuracy of the navigation information are ensured, further improving the operating efficiency of the scenic area car and the tourist experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 A schematic diagram of a flow chart of a method for establishing a smart path for a scenic area vehicle based on an AI algorithm according to an embodiment of the present invention;

[0058] Figure 2This is a structural diagram of a scenic area car intelligent path planning system based on an AI algorithm according to an embodiment of the present invention. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0060] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0061] Figure 1 FIG. 1 is a flow chart of a method for establishing a smart path for a scenic spot car based on an AI algorithm according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0062] S101. Collect geographic information data of all roads in the scenic area, and convert the geographic information data into a digital road network topology structure; perform feature extraction on the digital road network topology structure based on a deep learning neural network, and extract road width parameters, road slope parameters, road curvature parameters, and road condition parameters; input the road width parameters, road slope parameters, road curvature parameters, and road condition parameters into a preset road network evaluation model, generate a scenic area road capacity score, and establish a weighted road network model including the road capacity score;

[0063] S102. Collect real-time GPS positioning data and vehicle-mounted sensor data of all vehicles in the scenic area, obtain the current position information, driving speed information and driving direction information of the vehicle based on the GPS positioning data; use the vehicle-mounted sensor data to obtain obstacle distribution information and road congestion information around the vehicle; input the current position information, driving speed information, driving direction information, obstacle distribution information and road congestion information into a preset reinforcement learning model, and combine with the weighted road network model to generate a real-time optimal path planning strategy;

[0064] S103. Based on the real-time optimal path planning strategy, the initial recommended path is calculated; the changes in road conditions on the initial recommended path are monitored in real time, and when changes in road conditions are detected, the changed road condition data is re-input into the reinforcement learning model; the reinforcement learning model evaluates and dynamically adjusts the current path in real time based on the path reward and punishment mechanism, outputs an updated optimal path plan, and sends the updated optimal path plan to the navigation system of the scenic area car for path navigation.

[0065] The specific implementation method of the scenic area car intelligent path establishment planning method based on AI algorithm is as follows:

[0066] First, the geographic information data of all roads in the scenic area is collected. This includes using high-precision GPS equipment and laser radar scanners to comprehensively map the scenic area roads and obtain information such as the longitude and latitude coordinates, altitude, and width of the roads. At the same time, drone aerial photography is used to obtain overhead images of the scenic area roads. These original geographic data are imported into the GIS system to construct a vector electronic map of the scenic area roads, forming a digital road network topology containing nodes and edges.

[0067] Next, a convolutional neural network is used to extract features from the topological structure of the digital road network. Specifically, the road network data is converted into an image format and input into a pre-trained ResNet-50 network to extract high-level semantic features of the road. Then, a fully connected layer is used to map the features to specific road parameters, including road width, slope, curvature of turns, and road conditions. For example, the road width parameter ranges from 0 to 20 meters, the slope parameter ranges from -45° to 45°, the curvature parameter ranges from 0 to 1, and the road condition parameter ranges from 0 to 100 (0 indicates extremely poor road conditions, and 100 indicates excellent road conditions).

[0068] The extracted road parameters are input into the pre-trained random forest model, which uses the road parameters as input features and outputs a road capacity score of 0-100. The higher the score, the stronger the road capacity. For example, a road with a width of 10 meters, a slope of 5°, a turning curvature of 0.2, and a road condition parameter of 90 may have a capacity score of 85. Based on the scoring results, a weighted road network model is established, and the weight of the road section is proportional to its capacity score.

[0069] Then, the GPS positioning data of all cars in the scenic area is collected in real time to obtain the latitude and longitude coordinates, instantaneous speed and driving direction of each car. At the same time, the data of sensors such as on-board cameras and lidar are collected to identify the distribution of obstacles within 20 meters around the vehicle. In addition, based on the vehicle GPS trajectory data, the real-time average driving speed of each section of the road is calculated to evaluate the degree of road congestion. For example, if the normal driving speed of a section of road is 40km / h and the current average speed is only 10km / h, the section is judged to be severely congested.

[0070] The position, speed, direction information of the car, as well as the distribution of surrounding obstacles and the degree of road congestion are input into the pre-trained deep Q network model. The reinforcement learning model takes the current state as input and outputs the optimal action sequence to the destination. When training the model, the reward function is set as: +100 points for reaching the destination, -100 points for colliding with obstacles, and -1 minute / second for driving in congested sections. The actions output by the model include: going straight, turning left, turning right, slowing down, stopping, etc. Combined with the weighted road network model, a real-time optimal path planning strategy is generated that takes into account factors such as road conditions and congestion.

[0071] Based on the optimal path planning strategy, the A* algorithm is used to calculate the initial recommended path. For example, the recommended path from the entrance of the scenic area to a certain scenic spot is: go straight for 500 meters, turn right into the mountain road, and drive uphill for 2 kilometers to reach the destination. At the same time, the road conditions on the recommended path are continuously monitored. When a change in road conditions is detected, such as a sudden traffic accident causing a blockage on a certain road section, the updated road condition data is re-input into the reinforcement learning model.

[0072] The reinforcement learning model evaluates the current path in real time based on the path reward and punishment mechanism. For example, negative rewards are given when encountering a blocked section of road, and positive rewards are given when a new unblocked shortcut is found. The model dynamically adjusts the path planning strategy through continuous trial and error and learning, and outputs the updated optimal path plan. Finally, the updated path plan is sent to the navigation system of the scenic area car in the form of voice broadcast and map visualization to guide the driver to navigate in real time.

[0073] The beneficial effects of this method include:

[0074] 1. Through deep learning and reinforcement learning algorithms, intelligent modeling and dynamic path planning of complex scenic road networks are realized, which greatly improves the accuracy and real-time performance of navigation.

[0075] 2. It takes into account multi-dimensional parameters such as road width, slope, curvature, as well as real-time road conditions and vehicle status, making the planned path safer and more reliable, and improving the operating efficiency of scenic area vehicles.

[0076] 3. It has the ability to monitor road conditions in real time and adjust routes dynamically, and can effectively respond to emergencies, thus improving the robustness and adaptability of the system and providing tourists with a better riding experience.

[0077] In an optional implementation, the method further includes:

[0078] Road width information is collected through lidar point cloud data, and the RANSAC algorithm is used to extract the road surface boundary and calculate the width parameters. Sampling points are set every five meters on the road, and the sliding window method is used to perform local averaging on the sampling points to obtain a road width parameter matrix. The cubic spline interpolation method is used to construct a continuous slope curve based on elevation data, and equally spaced sampling is performed on the continuous slope curve to obtain discrete slope values ​​to form a road slope parameter matrix. The center line of the road is used to fit the arc through the least squares method to calculate the curvature radius, and the road curvature parameter matrix is ​​obtained.

[0079] This method collects road width information through LiDAR point cloud data. First, a high-precision LiDAR scanner is used to scan the road in all directions to obtain three-dimensional point cloud data including the road surface and the surrounding environment. The point cloud data contains the spatial coordinate information (x, y, z) and reflection intensity value of each sampling point.

[0080] Next, the RANSAC (Random Consistent Sampling) algorithm is used to extract the road surface boundary from the point cloud data. The RANSAC algorithm randomly selects sample points from the point cloud in an iterative manner, fits the plane model, and calculates the number of inliers that conform to the model. After multiple iterations, the plane model with the largest number of inliers is finally selected as the road surface. Points that do not conform to the road surface plane model are eliminated, and the remaining point set is the road surface point cloud.

[0081] In the road point cloud, the density clustering method is used to identify the curbstones on the left and right sides as the road boundary. A sampling point is set every 5 meters along the center line of the road. At each sampling point, the horizontal distance between the left and right curbstones is calculated as the road width at that location.

[0082] In order to eliminate the influence of local noise, the width values ​​obtained by sampling are smoothed by sliding windows. The window size is selected as 5 sampling points, and the width values ​​in the window are averaged to obtain the smoothed width values. The smoothed width values ​​are arranged in the order of sampling points to form a road width parameter matrix.

[0083] For example, a section of road is 100 meters long, with a sampling point every 5 meters, for a total of 20 sampling points. The width value (unit: meter) obtained by the original sampling is:

[0084] [7.2,7.1,7.3,7.0,7.2,7.1,7.4,7.2,7.3,7.1,7.0,7.2,7.3,7.1,7.2,7.4,7.3,7.2,7.1,7.3];

[0085] After averaging the 5-point sliding window, the smoothed width parameter matrix is:

[0086] [7.16,7.14,7.18,7.16,7.20,7.22,7.20,7.18,7.14,7.16,7.16,7.18,7.22,7.24,7.22,7.20];

[0087] Next, a continuous slope curve is constructed based on the elevation data using the cubic spline interpolation method. First, the elevation information on the road centerline is extracted from the LiDAR point cloud data to obtain a series of discrete (distance, elevation) data points. The cubic spline function is selected as the interpolation basis function, and a cubic polynomial curve segment is constructed between adjacent data points to ensure that the first-order and second-order derivatives are continuous at the nodes.

[0088] Using the constructed cubic spline function, the elevation value of any position on the center line of the road can be obtained. Sampling is performed at equal intervals throughout the entire road range, and the ratio of the elevation difference between adjacent sampling points to the horizontal distance is calculated, which is the slope value at that location. The discrete slope values ​​obtained by sampling are arranged in order to form a road slope parameter matrix.

[0089] For example, a section of road is 100 meters long and a sampling point is taken every 10 meters. The slope value (unit: %) calculated is:

[0090] [2.5,3.1,3.8,4.2,3.9,3.5,3.0,2.8,2.6,2.4];

[0091] Finally, the curvature radius is calculated by fitting the arc using the least squares method using the road centerline, and the road curvature parameter matrix is ​​obtained. First, the plane coordinate sequence of the road centerline is extracted from the lidar point cloud data. A road section of a certain length (such as 50 meters) is selected and the arc is fitted using the least squares method.

[0092] The specific steps are as follows:

[0093] 1) Select the center line coordinate point (xi,yi) within the road section, i=1,2,...,n.

[0094] 2) Construct the objective function to minimize the sum of the squares of the distances from all points to the center of the circle.

[0095] 3) Solve the extreme point of the objective function and obtain the center coordinates (a, b) and radius R of the best fitting circle.

[0096] 4) Calculate the curvature value k=1 / R of the road section.

[0097] Slide the fitting window along the road direction and repeat the above steps to obtain the curvature value of each road section. Arrange these curvature values ​​in order to form a road curvature parameter matrix.

[0098] For example, a section of road is 200 meters long, and the curvature is calculated every 50 meters. The curvature value (unit: 1 / km) is:

[0099] [0.5,0.8,1.2,1.5];

[0100] The beneficial effects of this method include:

[0101] 1. Improved the accuracy and efficiency of road geometry parameter extraction. High-precision three-dimensional information is obtained through lidar point cloud data. Combined with RANSAC algorithm, density clustering and other methods, it can accurately identify road boundaries and extract road width information. The sliding window average processing is used to effectively eliminate the influence of local noise and improve the reliability of width parameters.

[0102] 2. The continuous expression of road slope and curvature is realized. The continuous slope curve is constructed by using cubic spline interpolation method, which overcomes the limitations of traditional discrete sampling method and can obtain slope information at any position. The curvature is calculated by fitting the arc with the least square method, which ensures the smoothness and continuity of curvature calculation.

[0103] 3. A standardized road geometry parameter matrix is ​​formed, which provides an important basis for subsequent road design, construction and maintenance. Key geometric parameters such as width, slope and curvature are stored in matrix form, which is convenient for data analysis and processing. The standardized parameter expression method is conducive to the horizontal comparison between different road sections and provides support for the overall planning and optimization of the road network.

[0104] In an optional implementation, the road width parameter, road slope parameter, road curvature parameter and road condition parameter are input into a preset road network evaluation model to generate a scenic area road capacity score, and establishing a weighted road network model including a road capacity score includes:

[0105] The ground image recognition results are combined with the vehicle-mounted sensor data to construct a road condition assessment index matrix including road surface smoothness parameters, friction coefficient parameters and damage degree parameters;

[0106] A width evaluation subnetwork is established for the road width parameter matrix, a slope evaluation subnetwork is established for the road slope parameter matrix, a curvature evaluation subnetwork is established for the road curvature parameter matrix, and a road condition evaluation subnetwork is established for the road condition evaluation index matrix; the evaluation subnetworks all use a soft attention mechanism to dynamically adjust feature weights, and achieve synergy of multiple key features through a multi-head attention mechanism;

[0107] The feature weights output by the evaluation subnetwork are weighted and summed with the corresponding standardized feature values ​​and environmental correction coefficients to obtain a road capacity score, wherein the environmental correction coefficient is updated in real time according to weather conditions and time period information through a Bayesian optimization method; the road capacity score adopts a standardized scale of 0-100;

[0108] A weighted road network graph structure is constructed based on the road capacity score, the road network node information is constructed as a vertex set, the road section information is constructed as an edge set, and the weight information based on the capacity score is constructed as an edge weight matrix; the edge weight matrix adopts a time-varying weight mechanism to dynamically adjust the basic score through time factors and weather factors; an exponential decay function is introduced in the edge weight calculation to simulate the cumulative effect of long-distance travel.

[0109] The specific implementation is as follows:

[0110] First, the basic data of scenic area roads are obtained, including road width parameters, road slope parameters, road curvature parameters and road condition parameters. The road width parameters are measured by a high-precision laser scanner to form a width parameter matrix. The road slope parameters are measured along the route using an inclination sensor to obtain the slope value and construct a slope parameter matrix. The road curvature parameters are measured by the vehicle-mounted inertial measurement unit to obtain the curvature value of each road section and generate a curvature parameter matrix.

[0111] Next, the road condition data is obtained using ground image recognition technology and vehicle-mounted sensors. The ground image acquisition vehicle drives along the route, and the high-resolution camera on board takes road images, and uses deep learning algorithms to identify road damage, cracks, and other conditions. At the same time, the on-board laser scanner measures the road surface flatness, and the tire sensor measures the road surface friction coefficient. These data are integrated to form a road condition evaluation index matrix, which includes road surface flatness parameters, friction coefficient parameters, and damage degree parameters.

[0112] Then, for different types of parameter matrices, evaluation subnetworks are constructed respectively. For the road width parameter matrix, a width evaluation subnetwork is established, and a convolutional neural network is used to extract width features. For the road slope parameter matrix, a slope evaluation subnetwork is constructed, and a recurrent neural network is used to capture the slope change trend. For the road curvature parameter matrix, a curvature evaluation subnetwork is designed, and a graph convolutional network is used to analyze the curvature distribution. For the road condition evaluation index matrix, a road condition evaluation subnetwork is established, and a multimodal fusion network is used to integrate various indicators.

[0113] In each evaluation subnetwork, a soft attention mechanism is introduced to achieve dynamic adjustment of feature weights. Specifically, the attention score of each feature is calculated, the score is normalized and multiplied by the feature to obtain the weighted feature. At the same time, a multi-head attention mechanism is adopted, multiple attention heads are set, each head calculates the attention weight independently, and finally the results of multiple heads are spliced. This allows the network to pay attention to multiple important features and improve the expressiveness of the model.

[0114] The feature weights output by the evaluation subnetwork are multiplied by the standardized feature values ​​and then multiplied by the environmental correction coefficient to obtain the evaluation scores. The environmental correction coefficient is updated in real time through the Bayesian optimization method, taking into account the current weather conditions and time period information. For example, the weight of the road condition score is reduced on rainy days, and the weight of the width score is increased at night. The weighted sum of the scores is mapped to a standardized scale of 0-100 to obtain the final road capacity score.

[0115] Based on the road capacity score, a weighted road network graph structure is constructed. The key locations such as intersections and turning points in the road network are taken as vertices to form a vertex set. The road sections connecting the vertices are taken as edges to form an edge set. The road capacity score is used as the weight of the edge to generate an edge weight matrix. A time-varying weight mechanism is adopted to introduce time factors and weather factors to dynamically adjust the weights. For example, the weight of congested sections is increased during peak hours in the morning and evening, and the weight of all edges is reduced in rainy and snowy weather.

[0116] In the edge weight calculation, an exponential decay function is introduced to simulate the cumulative effect of long-distance travel. Specifically, a decay coefficient is set. The longer the edge is, the more obvious the cumulative effect is, and the final weight increases exponentially. This can more accurately reflect the actual difficulty of long-distance travel.

[0117] Finally, based on the constructed weighted road network model, applications such as shortest path planning and traffic flow prediction can be performed. For example, the Dijkstra algorithm can be used to calculate the optimal path between two points, and the congestion level of each road section can be predicted based on the edge weights.

[0118] The beneficial effects of this technical solution include:

[0119] 1. By integrating multi-source data and comprehensively considering factors such as road width, slope, curvature and road conditions, a comprehensive assessment of the road capacity in scenic areas is achieved, which improves the accuracy and reliability of the assessment.

[0120] 2. The introduction of soft attention mechanism and multi-head attention mechanism realizes the dynamic adjustment of feature weights and the synergy of multiple features, enhancing the adaptability and expression ability of the model.

[0121] 3. The time-varying weight mechanism and exponential decay function are adopted, taking into account dynamic factors such as time and weather as well as the cumulative effect of long-distance travel, making the constructed weighted road network model closer to actual traffic conditions and providing reliable data support for subsequent path planning and traffic management.

[0122] In an optional implementation, the current location information, driving speed information, driving direction information, obstacle distribution information and road congestion information are input into a preset reinforcement learning model, combined with the weighted road network model, to generate a real-time optimal path planning strategy, including:

[0123] The current position information of the vehicle is obtained through a positioning sensor, the driving speed information of the vehicle is obtained through a vehicle-mounted speed sensor, the driving direction information of the vehicle is obtained through a vehicle-mounted direction sensor, the obstacle distribution information is obtained through a vehicle-mounted laser radar scan, and the road congestion degree information is obtained through a road network monitoring device; the current position information, the driving speed information, and the driving direction information are mapped to a state space, and the obstacle distribution information and the road congestion degree information are mapped to an environment space;

[0124] Construct a dual Q learning network structure, including an evaluation network and a target network. The evaluation network inputs state space information and environment space information and outputs action value evaluation results. The target network calculates the target Q value based on historical evaluation data. Use a priority experience replay mechanism to store and sample training data, determine the priority of training samples according to the time difference error, and give priority to samples with larger time difference error for network update.

[0125] The action value evaluation result is integrated with the pre-built weighted road network model, and the path decision is sampled and evaluated based on the Monte Carlo tree search method. During the search process, the search tree is pruned and optimized in combination with the target Q value; a comprehensive score is calculated for each candidate path, and the comprehensive score includes three evaluation dimensions: path travel time, path safety, and path stability;

[0126] An optimal decision path is selected based on the comprehensive score, and a progressive dynamic programming method is used to locally optimize the optimal decision path. During the optimization process, vehicle steering constraints and acceleration and deceleration constraints are considered to ensure that the planned path meets the vehicle dynamics characteristics. The progressive dynamic programming method optimizes the path in segments through a rolling time domain to achieve real-time response to dynamic changes in the environment.

[0127] This embodiment provides a real-time optimal path planning method based on reinforcement learning. The method first obtains the current state information and environmental information of the vehicle through a variety of vehicle-mounted sensors and road network monitoring equipment, then inputs this information into a preset reinforcement learning model, and generates a real-time optimal path planning strategy in combination with a weighted road network model. The specific implementation steps are as follows:

[0128] First, the vehicle status information is obtained through the on-board sensors. The positioning sensor obtains the current latitude and longitude coordinates of the vehicle, such as (116.404, 39.915); the speed sensor obtains the current speed of the vehicle, such as 60km / h; the direction sensor obtains the direction of the vehicle, such as the north direction. At the same time, the on-board laser radar obtains the distribution information of surrounding obstacles through 360-degree scanning, such as a stationary vehicle 10 meters ahead and a pedestrian 5 meters to the left. In addition, the road network monitoring equipment obtains the current road congestion level information, such as the congestion index of the road ahead is 0.8.

[0129] Next, the acquired information is mapped to the state space and the environment space. The vehicle position, speed, and direction information are normalized and mapped to the state space in the interval [-1,1]. The obstacle distribution information is converted into an occupancy gridmap, and the road congestion information is converted into a heat map, which together constitute the environment space.

[0130] Then, a dual Q-learning network structure is constructed. The evaluation network uses a 3-layer fully connected neural network. The number of neurons in the input layer is the sum of the state space dimension and the environment space dimension, the number of neurons in the hidden layer is 256, and the number of neurons in the output layer is the action space dimension. The target network structure is the same as the evaluation network, but the parameter update frequency is lower. The priority experience replay mechanism uses the SumTree data structure to store samples, and calculates the sample priority p=|e|+ε according to the time difference error e, where ε is a small positive number.

[0131] The action value output by the evaluation network is integrated with the weighted road network model. The weight of each road in the weighted road network model is determined by factors such as length, speed limit, and number of lanes. The Upper Confidence Bounds algorithm is used for Monte Carlo tree search with a search depth of 20. The search tree is α-β pruned using the target Q value at each decision node. The comprehensive score S = w1*T+w2*Sa+w3*St is calculated for each candidate path, where T is the travel time, Sa is the safety score, St is the stability score, and w1, w2, and w3 are weight coefficients.

[0132] The optimal decision path is selected according to the comprehensive score, and the rolling horizon dynamic programming method is used for local optimization. The planning time domain is divided into 5 time steps, each time step corresponds to 2 seconds. In each time step, the dynamic programming algorithm is used to optimize the local path considering the constraints of the vehicle's maximum steering angle of 30° and the maximum acceleration of 2m / s2. The optimization goal is to minimize the weighted sum of the path length and the congestion degree.

[0133] Finally, the optimized real-time optimal path planning strategy is output, including the vehicle trajectory point sequence, speed planning curve and steering angle planning curve in the next 10 seconds. The system re-executes the planning process every 1 second to achieve real-time response to dynamic changes in the environment.

[0134] The beneficial effects of this method include:

[0135] 1) By integrating multi-source sensor data and reinforcement learning models, accurate perception and rapid decision-making of complex dynamic environments are achieved, and the real-time and robustness of path planning are improved.

[0136] 2) The use of dual Q-learning networks and priority experience replay mechanism effectively improves the convergence speed and generalization ability of the reinforcement learning model, enabling the model to cope with various complex traffic scenarios.

[0137] 3) Combining Monte Carlo tree search and dynamic programming algorithms, local dynamic optimization is taken into account while ensuring global optimality, which not only ensures the overall optimality of the path, but also satisfies the vehicle dynamics constraints and improves the feasibility of the planned path.

[0138] In an optional implementation, an optimal decision path is selected based on the comprehensive score, and a progressive dynamic programming method is used to locally optimize the optimal decision path. During the optimization process, vehicle steering constraints and acceleration and deceleration constraints are considered to ensure that the planned path meets the vehicle dynamics characteristics; the progressive dynamic programming method optimizes the path in segments through a rolling time domain to achieve real-time response to dynamic changes in the environment, including:

[0139] Establish a vehicle dynamics model, including vehicle center of mass motion state parameters, steering system parameters and power system parameters. Collect vehicle lateral acceleration, longitudinal acceleration, sideslip angular velocity, steering angular velocity and wheel speed data in real time through on-board sensors, and perform online identification and update of dynamics model parameters based on the extended Kalman filter algorithm.

[0140] Construct a path optimization objective function, and set path length minimization, steering energy minimization, and speed fluctuation minimization as optimization sub-objectives. At the same time, introduce vehicle steering angle constraints, steering angular velocity constraints, longitudinal acceleration constraints, and lateral acceleration constraints as hard constraints, and set the safe distance between the road boundary and the obstacle as a soft constraint.

[0141] The global planning path is segmented, and the optimization interval is defined by the sliding time window method. The optimization interval is divided into multiple equal time steps. The path curve is parameterized by a quintic polynomial in each time step to construct a state variable sequence including position, velocity, and acceleration.

[0142] For each optimization interval, a state transfer network is constructed based on the dynamic programming algorithm. The network nodes represent the feasible states of the vehicle, and the network edges represent the state transfer costs. During the state transfer process, the feasibility of the state transfer is verified through the vehicle dynamics model, and the state transfer edges that do not meet the constraints are eliminated.

[0143] A hierarchical search strategy is used to find the optimal path in the state transition network. First, a coarse-grained search is performed to determine the approximate search direction, and then a fine-grained search is performed near the optimal path to optimize the local trajectory. A heuristic evaluation function is introduced during the search process, combining the attraction of the target point with the repulsion of the obstacle as the evaluation basis.

[0144] The optimization interval is rolled and updated based on the progressive mechanism. The previous part of the current optimization interval is retained as the execution trajectory. The optimization window is slid forward by a fixed step size, and the state transfer network construction and path search process are repeated. During the interval update process, the path characteristics of adjacent optimization intervals are ensured to be continuous.

[0145] After selecting the optimal decision path based on the comprehensive score, the progressive dynamic programming method is used to locally optimize the optimal decision path. First, a vehicle dynamics model is established, including the vehicle center of mass motion state parameters, steering system parameters and power system parameters. The vehicle lateral acceleration, longitudinal acceleration, sideslip angular velocity, steering angular velocity and wheel speed data are collected in real time through on-board sensors, and the extended Kalman filter algorithm is used to identify and update the dynamic model parameters online. For example, the initial parameters such as vehicle mass of 1500kg, wheelbase of 2.8m, front and rear track widths of 1.6m and 1.58m respectively, and tire radius of 0.3m can be set, and then these parameters can be dynamically adjusted based on the real-time collected sensor data.

[0146] Next, we construct the path optimization objective function, and minimize the path length, steering energy, and speed fluctuation as optimization sub-objectives. At the same time, we introduce the vehicle steering angle constraint, steering angular velocity constraint, longitudinal acceleration constraint, and lateral acceleration constraint as hard constraints, and the road boundary and obstacle safety distance as soft constraints. For example, we can set the maximum steering angle to 35°, the maximum steering angular velocity to 15° / s, the maximum longitudinal acceleration to 2m / s^2, the maximum lateral acceleration to 4m / s^2, and the minimum safety distance from obstacles to 1.5m.

[0147] The global planning path is segmented, and the optimization interval is defined by the sliding time window method. The optimization interval is divided into multiple equal time steps. The path curve is parameterized by a quintic polynomial in each time step to construct a state variable sequence containing position, velocity, and acceleration. Specifically, the sliding time window can be set to 5s and the time step to 0.1s, then each optimization interval contains 50 discrete time points. For each time point, the position, velocity, and acceleration of the vehicle in the x and y directions are represented by a quintic polynomial.

[0148] For each optimization interval, a state transition network is constructed based on the dynamic programming algorithm. The network nodes represent the feasible states of the vehicle, and the network edges represent the state transition costs. During the state transition process, the feasibility of the state transition is verified by the vehicle dynamics model, and the state transition edges that do not meet the constraints are eliminated. For example, the lateral position of the vehicle can be discretized into 21 values, the longitudinal position can be discretized into 11 values, and the speed can be discretized into 7 values. Then the state space size of each time step is 1617 nodes. For each pair of nodes between adjacent time steps, the state transition cost is calculated and it is verified whether the dynamic constraints are met, and a feasible state transition edge is established.

[0149] A hierarchical search strategy is used to find the optimal path in the state transition network. First, a coarse-grained search is performed to determine the approximate search direction, and then a fine-grained search is performed near the optimal path to optimize the local trajectory. A heuristic evaluation function is introduced in the search process, combining the attraction of the target point with the repulsion of the obstacle as the evaluation basis. Specifically, the state space can be coarsened into a 5×3×3 grid for search to determine the approximate path direction, and then two grids are expanded on both sides of the optimal rough path for fine-grained search. The heuristic function can be set as a weighted sum of 80% of the distance to the target point and 20% of the inverse of the distance to the nearest obstacle.

[0150] Based on the progressive mechanism, the optimization interval is updated in a rolling manner, the previous part of the current optimization interval is retained as the execution trajectory, the optimization window is slid forward by a fixed step length, and the state transition network construction and path search process are repeated. During the interval update process, the path characteristics of adjacent optimization intervals are ensured to be continuous. For example, the first 2s trajectory of the current optimization interval can be retained as the execution trajectory, and then the optimization window is slid forward by 2s to ensure that the new and old optimization intervals overlap by 3s. The new optimization results are smoothed in the overlapping interval to ensure the continuity of the path curvature.

[0151] The beneficial effects of this method include: 1) by combining the dynamic programming algorithm with the rolling optimization strategy, it can ensure global optimality while having the ability to respond quickly to environmental changes; 2) by adopting a multi-level search strategy and heuristic evaluation function, it can reduce the computational complexity while ensuring the search efficiency; 3) by taking into account the vehicle dynamics constraints, it can ensure that the planned path meets the vehicle's motion characteristics, thereby improving trajectory tracking accuracy and driving comfort.

[0152] In an optional implementation, the reinforcement learning model evaluates and dynamically adjusts the current path in real time based on the path reward and punishment mechanism, outputs an updated optimal path plan, and sends the updated optimal path plan to the navigation system of the scenic spot car for path navigation, including:

[0153] Construct a multi-level path reward and punishment evaluation framework, divide the path evaluation indicators into safety indicators, efficiency indicators and comfort indicators, and calculate the scores of various evaluation indicators based on the driving data collected in real time by the on-board sensors; the safety indicators include the minimum distance to obstacles, the minimum distance to other vehicles, and the road curvature adaptability; the efficiency indicators include the average driving speed, the path length ratio, and the degree of congestion and detour; the comfort indicators include the acceleration fluctuation, the steering angle change rate, and the road surface smoothness;

[0154] A hierarchical reinforcement learning architecture is used to dynamically evaluate the path. The top-level policy network determines the path adjustment direction based on the global path characteristics, and the bottom-level execution network generates specific adjustment strategies based on local road condition information. In the evaluation process, spatiotemporal attenuation factors are integrated, and differentiated weights are given to historical evaluation data to highlight the impact of recent road condition changes.

[0155] Construct an adaptive reward and penalty function, where the reward items include the reduction in path travel time, the improvement in safety margin, and the improvement in comfort, and the penalty items include the degree of deviation from the original path, the loss of crossing congested areas, and the loss of frequent lane changes; the reward and penalty function adopts a dynamic weight mechanism to adaptively adjust the weight coefficients of each score according to the characteristics of the driving scene;

[0156] Based on the evaluation results of the reward and punishment function, the deep deterministic policy gradient algorithm is used to optimize the path plan, the path adjustment strategy is generated through the Actor network, and the strategy value is evaluated through the Critic network; the path smoothing constraint is introduced in the strategy generation process to ensure the continuity and feasibility of the path adjustment;

[0157] The optimized path plan is converted into a navigation instruction sequence, including the expected heading angle, expected speed, turning prompts and distance prompts, and the navigation instruction sequence is sent to the navigation system of the scenic area car through the on-board communication module; the navigation system performs real-time navigation based on the received navigation instruction sequence, and transmits the feedback information during the navigation process back to the path optimization module, forming a closed-loop optimization mechanism.

[0158] In this implementation, the reinforcement learning model evaluates and dynamically adjusts the current path in real time based on the path reward and punishment mechanism, outputs the updated optimal path plan, and sends the updated optimal path plan to the navigation system of the scenic area car for path navigation. The specific implementation process is as follows:

[0159] First, a multi-level path reward and punishment evaluation framework is constructed. The path evaluation indicators are divided into three levels: safety indicators, efficiency indicators, and comfort indicators. The safety indicator includes three sub-indicators: minimum distance to obstacles, minimum distance to other vehicles, and road curvature adaptability. The efficiency indicator includes three sub-indicators: average driving speed, path length ratio, and congestion detour degree. The comfort indicator includes three sub-indicators: acceleration fluctuation, steering angle change rate, and road surface smoothness.

[0160] Taking the safety index as an example, the minimum distance to obstacles is obtained through real-time scanning by the on-board laser radar, and the safety threshold is set to 3 meters. The minimum distance to other vehicles is obtained through on-board millimeter-wave radar detection, and the safety threshold is set to 5 meters. The road curvature adaptability is calculated based on the lateral acceleration data obtained by the on-board inertial measurement unit, and the threshold is set to 0.4g. For each sub-indicator, if it is higher than the safety threshold, the score is 1, and if it is lower than the threshold, the score is converted proportionally. Finally, the average score of the three sub-indicators is taken as the comprehensive score of the safety index.

[0161] The calculation methods of efficiency index and comfort index are similar. For example, the average driving speed is calculated by the location data collected by the vehicle GPS module, and compared with the maximum speed limit of 40km / h stipulated in the scenic area to calculate the score. The path length ratio is determined by comparing the ratio of the current path length to the straight-line distance. The closer the ratio is to 1, the higher the score. The degree of congestion detour evaluates the rationality of the detour path based on real-time traffic information. The shorter the detour distance and the better the effect of avoiding congested sections, the higher the score.

[0162] Next, a hierarchical reinforcement learning architecture is used to dynamically evaluate the path. The top-level policy network determines the direction of path adjustment based on global path features, and the input features include the coordinates of the starting and ending points, the sequence of scenic spots along the way, and the topological structure of the road network. The bottom-level execution network generates specific adjustment strategies based on local road condition information, and the input features include the current location, the distribution of surrounding obstacles, the curvature of the road, the degree of congestion, etc. In the evaluation process, the spatiotemporal attenuation factors are integrated, and the historical evaluation data is given differentiated weights to highlight the impact of recent changes in road conditions. Specifically, the time attenuation coefficient is set to 0.9, the spatial attenuation coefficient is set to 0.95, the weight of historical data before time t is 0.9^t, and the weight of historical data d meters away from the current location is 0.95^d.

[0163] Then, an adaptive reward and penalty function is constructed. The reward items include the reduction in path travel time, the improvement in safety margin, and the improvement in comfort. For example, if the optimized path saves 5 minutes compared to the original path, a 50-point reward is given; if the minimum distance to the obstacle increases by 1 meter, a 20-point reward is given; if the acceleration fluctuation is reduced by 0.1g, a 10-point reward is given. The penalty items include the degree of deviation from the original path, the loss of crossing the congested area, and the loss of frequent lane changes. For example, if the deviation from the original path exceeds 500 meters, a -30-point penalty is given; if a congested section is crossed, a -20-point penalty is given; if the number of lane changes within 1 kilometer exceeds 3 times, a -10-point penalty is given. The reward and penalty function adopts a dynamic weight mechanism to adaptively adjust the weight coefficients of each score according to the characteristics of the driving scene. In congested sections, the weight of efficiency indicators is increased; in mountain roads with many bends, the weight of safety indicators is increased; on flat and open roads, the weight of comfort indicators is increased.

[0164] Based on the evaluation results of the reward and punishment function, a deep deterministic policy gradient algorithm is used to optimize the path plan. The path adjustment strategy is generated through the Actor network. The Actor network adopts a 4-layer fully connected neural network structure. The number of neurons in the input layer is the current state dimension. The hidden layers are 256, 128, and 64 neurons respectively, and the output layer is the action space dimension. The strategy value is evaluated through the Critic network. The Critic network also adopts a 4-layer fully connected structure. The input layer is the sum of the state dimension and the action dimension. The hidden layer has the same structure as the Actor network, and the output layer is 1 neuron to represent the Q value. Path smoothing constraints are introduced in the strategy generation process to ensure the continuity and feasibility of path adjustment. In the specific implementation, the cubic spline interpolation algorithm is applied to the generated path point sequence to ensure the continuity of the path curvature.

[0165] Finally, the optimized path plan is converted into a navigation instruction sequence, including the expected heading angle, expected speed, steering prompts and distance prompts. For example, "Turn left 200 meters ahead and maintain 30km / h." The navigation instruction sequence is sent to the navigation system of the scenic area car through the vehicle communication module. The navigation system performs real-time navigation based on the received navigation instruction sequence, and transmits the feedback information during the navigation process back to the path optimization module to form a closed-loop optimization mechanism. The feedback information includes actual driving trajectory, average speed, acceleration and other data, which are used to update the state space of the reinforcement learning model and continuously optimize the path planning strategy.

[0166] This technical solution has the following beneficial effects:

[0167] 1. Improved safety: Comprehensively consider safety factors through a multi-level evaluation framework, perceive the surrounding environment in real time, dynamically adjust the path to maintain a safe distance, and effectively reduce the risk of accidents.

[0168] 2. Efficiency optimization: Optimize routes based on real-time traffic information, intelligently avoid congested areas, select the best route, and significantly improve the operating efficiency of scenic area vehicles.

[0169] 3. Comfort enhancement: Comfort indicators such as acceleration fluctuation and steering smoothness are taken into consideration, and path smoothness constraints are introduced to ensure that the generated path is continuous and smooth, thereby improving passenger riding comfort.

[0170] Figure 2 FIG. 1 is a schematic diagram of a system for establishing a smart path planning system for a scenic area vehicle based on an AI algorithm according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0171] The first unit is used to collect geographic information data of all roads in the scenic area, convert the geographic information data into a digital road network topology structure; perform feature extraction on the digital road network topology structure based on a deep learning neural network, extract road width parameters, road slope parameters, road curvature parameters and road condition parameters; input the road width parameters, road slope parameters, road curvature parameters and road condition parameters into a preset road network evaluation model, generate a scenic area road capacity score, and establish a weighted road network model including the road capacity score;

[0172] The second unit is used to collect real-time GPS positioning data and vehicle-mounted sensor data of all vehicles in the scenic area, obtain the current position information, driving speed information and driving direction information of the vehicle based on the GPS positioning data; use the vehicle-mounted sensor data to obtain obstacle distribution information and road congestion information around the vehicle; input the current position information, driving speed information, driving direction information, obstacle distribution information and road congestion information into a preset reinforcement learning model, and combine with the weighted road network model to generate a real-time optimal path planning strategy;

[0173] The third unit is used to calculate the initial recommended path based on the real-time optimal path planning strategy; monitor the changes in road conditions on the initial recommended path in real time, and when a change in road conditions is detected, re-input the changed road condition data into the reinforcement learning model; the reinforcement learning model evaluates and dynamically adjusts the current path in real time based on the path reward and punishment mechanism, outputs an updated optimal path plan, and sends the updated optimal path plan to the navigation system of the scenic area car for path navigation.

[0174] According to a third aspect of the embodiments of the present invention,

[0175] An electronic device is provided, comprising:

[0176] processor;

[0177] a memory for storing processor-executable instructions;

[0178] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0179] A fourth aspect of the embodiments of the present invention is:

[0180] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0181] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0182] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. The intelligent path establishment and planning method for scenic area vehicles based on AI algorithm is characterized by: include: Collecting geographic information data of all roads in the scenic area and converting the geographic information data into a digital road network topology structure; Based on the deep learning neural network, the digital road network topology structure is feature extracted to extract road width parameters, road slope parameters, road curvature parameters and road condition parameters; the road width parameters, road slope parameters, road curvature parameters and road condition parameters are input into a preset road network evaluation model to generate a scenic area road capacity score, and establish a weighted road network model including the road capacity score; Collect real-time GPS positioning data and vehicle-mounted sensor data of all vehicles in the scenic area, and obtain the current location information, driving speed information and driving direction information of the vehicle based on the GPS positioning data; use the vehicle-mounted sensor data to obtain the obstacle distribution information and road congestion degree information around the vehicle; Input the current location information, driving speed information, driving direction information, obstacle distribution information and road congestion information into a preset reinforcement learning model, and combine with the weighted road network model to generate a real-time optimal path planning strategy; Based on the real-time optimal path planning strategy, an initial recommended path is calculated; changes in road conditions on the initial recommended path are monitored in real time, and when changes in road conditions are detected, the changed road condition data is re-input into the reinforcement learning model; The reinforcement learning model evaluates and dynamically adjusts the current path in real time based on the path reward and punishment mechanism, outputs an updated optimal path plan, and sends the updated optimal path plan to the navigation system of the scenic area car for path navigation.

2. The method according to claim 1, characterized in that The method further comprises: Road width information is collected through lidar point cloud data, and the RANSAC algorithm is used to extract the road surface boundary and calculate the width parameters. Sampling points are set every five meters on the road, and the sliding window method is used to perform local averaging on the sampling points to obtain a road width parameter matrix. The cubic spline interpolation method is used to construct a continuous slope curve based on elevation data, and equally spaced sampling is performed on the continuous slope curve to obtain discrete slope values ​​to form a road slope parameter matrix. The center line of the road is used to fit the arc through the least squares method to calculate the curvature radius, and the road curvature parameter matrix is ​​obtained.

3. The method according to claim 2, characterized in that Inputting the road width parameter, road slope parameter, road curvature parameter and road condition parameter into a preset road network evaluation model to generate a scenic area road capacity score, and establishing a weighted road network model including the road capacity score includes: The ground image recognition results are combined with the vehicle-mounted sensor data to construct a road condition assessment index matrix including road surface smoothness parameters, friction coefficient parameters and damage degree parameters; A width evaluation subnetwork is established for the road width parameter matrix, a slope evaluation subnetwork is established for the road slope parameter matrix, a curvature evaluation subnetwork is established for the road curvature parameter matrix, and a road condition evaluation subnetwork is established for the road condition evaluation index matrix; the evaluation subnetworks all use a soft attention mechanism to dynamically adjust feature weights, and achieve synergy of multiple key features through a multi-head attention mechanism; The feature weights output by the evaluation subnetwork are weighted and summed with the corresponding standardized feature values ​​and environmental correction coefficients to obtain a road capacity score, wherein the environmental correction coefficient is updated in real time according to weather conditions and time period information through a Bayesian optimization method; the road capacity score adopts a standardized scale of 0-100; A weighted road network graph structure is constructed based on the road capacity score, the road network node information is constructed as a vertex set, the road section information is constructed as an edge set, and the weight information based on the capacity score is constructed as an edge weight matrix; the edge weight matrix adopts a time-varying weight mechanism to dynamically adjust the basic score through time factors and weather factors; an exponential decay function is introduced in the edge weight calculation to simulate the cumulative effect of long-distance travel.

4. The method according to claim 1, characterized in that: Inputting the current location information, driving speed information, driving direction information, obstacle distribution information and road congestion information into a preset reinforcement learning model, and combining with the weighted road network model to generate a real-time optimal path planning strategy includes: The current position information of the vehicle is obtained through a positioning sensor, the driving speed information of the vehicle is obtained through a vehicle-mounted speed sensor, the driving direction information of the vehicle is obtained through a vehicle-mounted direction sensor, the obstacle distribution information is obtained through a vehicle-mounted laser radar scan, and the road congestion degree information is obtained through a road network monitoring device; the current position information, the driving speed information, and the driving direction information are mapped to a state space, and the obstacle distribution information and the road congestion degree information are mapped to an environment space; Construct a dual Q learning network structure, including an evaluation network and a target network. The evaluation network inputs state space information and environment space information and outputs action value evaluation results. The target network calculates the target Q value based on historical evaluation data. Use a priority experience replay mechanism to store and sample training data, determine the priority of training samples according to the time difference error, and give priority to samples with larger time difference error for network update. The action value evaluation result is integrated with the pre-built weighted road network model, and the path decision is sampled and evaluated based on the Monte Carlo tree search method. During the search process, the search tree is pruned and optimized in combination with the target Q value; a comprehensive score is calculated for each candidate path, and the comprehensive score includes three evaluation dimensions: path travel time, path safety, and path stability; An optimal decision path is selected based on the comprehensive score, and a progressive dynamic programming method is used to locally optimize the optimal decision path. During the optimization process, vehicle steering constraints and acceleration and deceleration constraints are considered to ensure that the planned path meets the vehicle dynamics characteristics. The progressive dynamic programming method optimizes the path in segments through a rolling time domain to achieve real-time response to dynamic changes in the environment.

5. The method according to claim 4, characterized in that The optimal decision path is selected based on the comprehensive score, and the optimal decision path is locally optimized by using a progressive dynamic programming method. During the optimization process, the vehicle steering constraint and the acceleration and deceleration constraint are considered to ensure that the planned path meets the vehicle dynamic characteristics. The progressive dynamic programming method optimizes the path in segments through a rolling time domain to achieve real-time response to dynamic changes in the environment, including: Establish a vehicle dynamics model, including vehicle center of mass motion state parameters, steering system parameters and power system parameters. Collect vehicle lateral acceleration, longitudinal acceleration, sideslip angular velocity, steering angular velocity and wheel speed data in real time through on-board sensors, and perform online identification and update of dynamics model parameters based on the extended Kalman filter algorithm. Construct a path optimization objective function, and set path length minimization, steering energy minimization, and speed fluctuation minimization as optimization sub-objectives. At the same time, introduce vehicle steering angle constraints, steering angular velocity constraints, longitudinal acceleration constraints, and lateral acceleration constraints as hard constraints, and set the safe distance between the road boundary and the obstacle as a soft constraint. The global planning path is segmented, and the optimization interval is defined by the sliding time window method. The optimization interval is divided into multiple equal time steps. The path curve is parameterized by a quintic polynomial in each time step to construct a state variable sequence including position, velocity, and acceleration. For each optimization interval, a state transfer network is constructed based on the dynamic programming algorithm. The network nodes represent the feasible states of the vehicle, and the network edges represent the state transfer costs. During the state transfer process, the feasibility of the state transfer is verified through the vehicle dynamics model, and the state transfer edges that do not meet the constraints are eliminated. A hierarchical search strategy is used to find the optimal path in the state transition network. First, a coarse-grained search is performed to determine the approximate search direction, and then a fine-grained search is performed near the optimal path to optimize the local trajectory. A heuristic evaluation function is introduced during the search process, combining the attraction of the target point with the repulsion of the obstacle as the evaluation basis. The optimization interval is rolled and updated based on the progressive mechanism. The previous part of the current optimization interval is retained as the execution trajectory. The optimization window is slid forward by a fixed step size, and the state transfer network construction and path search process are repeated. During the interval update process, the path characteristics of adjacent optimization intervals are ensured to be continuous.

6. The method according to claim 1, characterized in that The reinforcement learning model evaluates and dynamically adjusts the current path in real time based on the path reward and punishment mechanism, outputs an updated optimal path plan, and sends the updated optimal path plan to the navigation system of the scenic spot car for path navigation, including: Construct a multi-level path reward and punishment evaluation framework, divide the path evaluation indicators into safety indicators, efficiency indicators and comfort indicators, and calculate the scores of various evaluation indicators based on the driving data collected in real time by the on-board sensors; the safety indicators include the minimum distance to obstacles, the minimum distance to other vehicles, and the road curvature adaptability; the efficiency indicators include the average driving speed, the path length ratio, and the degree of congestion and detour; the comfort indicators include the acceleration fluctuation, the steering angle change rate, and the road surface smoothness; A hierarchical reinforcement learning architecture is used to dynamically evaluate the path. The top-level policy network determines the path adjustment direction based on the global path characteristics, and the bottom-level execution network generates specific adjustment strategies based on local road condition information. In the evaluation process, spatiotemporal attenuation factors are integrated, and differentiated weights are given to historical evaluation data to highlight the impact of recent road condition changes. Construct an adaptive reward and penalty function, where the reward items include the reduction in path travel time, the improvement in safety margin, and the improvement in comfort, and the penalty items include the degree of deviation from the original path, the loss of crossing congested areas, and the loss of frequent lane changes; the reward and penalty function adopts a dynamic weight mechanism to adaptively adjust the weight coefficients of each score according to the characteristics of the driving scene; Based on the evaluation results of the reward and punishment function, the deep deterministic policy gradient algorithm is used to optimize the path plan, the path adjustment strategy is generated through the Actor network, and the strategy value is evaluated through the Critic network; the path smoothing constraint is introduced in the strategy generation process to ensure the continuity and feasibility of the path adjustment; The optimized path plan is converted into a navigation instruction sequence, including the expected heading angle, expected speed, turning prompts and distance prompts, and the navigation instruction sequence is sent to the navigation system of the scenic area car through the on-board communication module; the navigation system performs real-time navigation based on the received navigation instruction sequence, and transmits the feedback information during the navigation process back to the path optimization module, forming a closed-loop optimization mechanism.

7. A scenic area car intelligent path establishment and planning system based on AI algorithm, used to implement the method described in any one of claims 1 to 6, characterized in that: include: The first unit is used to collect geographic information data of all roads in the scenic area and convert the geographic information data into a digital road network topology structure; Based on the deep learning neural network, the digital road network topology structure is feature extracted to extract road width parameters, road slope parameters, road curvature parameters and road condition parameters; the road width parameters, road slope parameters, road curvature parameters and road condition parameters are input into a preset road network evaluation model to generate a scenic area road capacity score, and establish a weighted road network model including the road capacity score; The second unit is used to collect real-time GPS positioning data and vehicle-mounted sensor data of all vehicles in the scenic area, and obtain the current position information, driving speed information and driving direction information of the vehicle based on the GPS positioning data; and use the vehicle-mounted sensor data to obtain the obstacle distribution information and road congestion degree information around the vehicle; Input the current location information, driving speed information, driving direction information, obstacle distribution information and road congestion information into a preset reinforcement learning model, and combine with the weighted road network model to generate a real-time optimal path planning strategy; A third unit is used to calculate an initial recommended path based on the real-time optimal path planning strategy; monitoring the changes of the road conditions on the initial recommended path in real time, and when a change in the road conditions is detected, re-inputting the changed road condition data into the reinforcement learning model; The reinforcement learning model evaluates and dynamically adjusts the current path in real time based on the path reward and punishment mechanism, outputs an updated optimal path plan, and sends the updated optimal path plan to the navigation system of the scenic area car for path navigation.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images