Unmanned vehicle local path planning method based on neural network and dynamic window method
Through the local path planning method of unmanned vehicles based on neural network and dynamic window method, the problem of traditional methods falling into local minimum value and insufficient generalization capabilities in complex environments is solved, and the goal reachability and Obstacle avoidance safety, possess independent learning ability and is suitable for different environments.
Patent Information
- Application Number
- CN202411270137.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-11
- Publication Date
- 2024-12-06
AI Technical Summary
Traditional local path planning methods are prone to falling into local minimum values in complex environments, lack generalization capabilities, and are difficult to directly deploy applications, and cannot effectively avoid obstacles.
The local path planning method of unmanned vehicles is adopted based on neural network and dynamic window method. By generating grid maps, building residual neural networks, iterative optimization, evaluating network performance, combining sensor information for path planning, and obtaining the most The best speed and angular speed are actually controlled.
It achieves the goal reachability and obstacle avoidance safety in complex environments, has the ability to learn independently, adapt to different environments, and is directly used for actual deployment.
Smart Images

Figure CN119088035A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of unmanned vehicle path planning, and in particular relates to an unmanned vehicle local path planning method based on a neural network and a dynamic window method. Background Art
[0002] Path planning is the core part of the motion planning system of unmanned vehicles. A reasonable path can ensure that the unmanned vehicle reaches the destination quickly and accurately while avoiding obstacles safely and effectively. It includes global path planning and local path planning. Global path planning is to plan a path in a known global environment, and local path planning is to plan a path in an unknown or dynamic environment. Since unmanned vehicles often cannot obtain global environmental information on the map, local path planning has a wider range of applications.
[0003] In traditional local path planning methods, when there are multiple obstacles or complex terrain, the algorithm may fall into a local minimum, making it impossible for the unmanned vehicle to reach its destination safely. In addition, traditional local path planning methods are usually designed for specific scenarios and lack the ability to generalize to different environments. When encountering unverified environments, they may not work effectively. There are already many feasible solutions for local path planning methods based on neural networks, but the planning result is often a trajectory, and a tracking algorithm is required to actually control the movement of the unmanned vehicle. There is no obstacle detection during the movement, so it is easy to encounter obstacles, and it is not convenient to directly deploy applications. Summary of the invention
[0004] In order to solve the above problems, the purpose of the present invention is to provide a local path planning method for an unmanned vehicle based on a neural network and a dynamic window method.
[0005] In order to achieve the above object, the unmanned vehicle local path planning method based on neural network and dynamic window method provided by the present invention comprises the following steps performed in sequence:
[0006] 1) Based on the working environment of the unmanned vehicle, the grid map is generated in batches and the A* algorithm is used to plan the global optimal path. Then, the global optimal path is sampled to obtain various local environmental conditions and corresponding azimuth actions, a data set is established, and the data set is divided into a training set and a test set in proportion;
[0007] 2) constructing a residual neural network, and then iteratively optimizing the residual neural network using the training data in the above training set to obtain a trained residual neural network;
[0008] 3) Using the test set to evaluate the above trained residual neural network, whether the trained residual neural network can achieve the expected effect, if the evaluation result is yes, the optimized residual neural network is obtained, otherwise return to step 1);
[0009] 4) Using the sensors on the unmanned vehicle to scan the current environment information to obtain the map information matrix, and then input the above-mentioned optimized residual neural network for path planning to obtain the corresponding planned azimuth action;
[0010] 5) Based on the above-mentioned planned azimuth action, the corresponding optimal angular velocity and optimal linear velocity are obtained by a dynamic window method, and then the unmanned vehicle moves to the next position according to the optimal angular velocity and optimal linear velocity;
[0011] 6) When the unmanned vehicle is moving, it continuously detects whether the target point appears in the detection window and there are no obstacles between the unmanned vehicle and the target point; if the detection result is yes, the navigation is considered to be completed, and the unmanned vehicle moves directly to the target point in a straight line and ends the navigation; otherwise, after the unmanned vehicle moves for a certain period of time at the optimal angular velocity and optimal linear velocity corresponding to the optimal trajectory, steps 4)-5) are repeated, and the next planning process is immediately started, and the vehicle moves along the optimal motion trajectory planned next time, and this process is repeated until the navigation is completed.
[0012] In step 1), based on the working environment of the unmanned vehicle, the grid map is generated in batches and the global optimal path is planned using the A* algorithm, and then the global optimal path is sampled to obtain various local environmental conditions and corresponding azimuth actions, a data set is established, and the specific method of dividing the data set into a training set and a test set in proportion is as follows:
[0013] 1.1) Select the working environment of the unmanned vehicle, build a map of the working environment, and extract geographical information including terrain and objects from the map;
[0014] 1.2) Input the above geographic information into map software such as ArcGIS and GIS to batch generate a large number of raster maps with similar geographic information, then randomly specify the starting point and target point in the raster map, and use the A* algorithm to plan the global optimal path;
[0015] 1.3) In the above global optimal path, a sampling point is set at a certain distance as the current sampling point, and the map information matrix within the 7×7 grid range around the current sampling point is collected as the local environment of the current sampling point;
[0016] 1.4) For each grid in the above map information matrix, if there is an obstacle in the grid, -1 is marked in the grid, otherwise the Euclidean distance from the center point of the grid to the target point is marked in the grid;
[0017] 1.5) Calculate the angle between the line connecting the next sampling point and the current sampling point and the horizontal direction and use it as the azimuth angle of the current sampling point;
[0018] 1.6) Taking the local environment of the current sampling point as input and the corresponding azimuth action as output, a complete set of training data is formed. Sampling all the global optimal paths in the batch-generated raster maps can obtain sufficient training data. All the training data constitute a data set, and the data set is divided into a training set and a test set at a ratio of 9:1.
[0019] In step 2), the residual neural network is constructed, and then the residual neural network is iteratively optimized using the training data in the training set to obtain the trained residual neural network. The specific method is as follows:
[0020] 2.1) constructing a residual neural network including an input layer, a pooling layer, a residual layer, an identity mapping layer and an output layer; the number of residual layers and identity mapping layers in the residual neural network is 20-30 layers, and the residual layer and the identity mapping layer are fully connected;
[0021] 2.2) Input the local environment conditions in the training set into the above residual neural network to obtain the output value M of the residual neural network;
[0022] 2.3) The mean square error loss function shown in Formula 1 is used to calculate the output value M of the residual neural network and the loss value Y of the corresponding azimuth action N in the training set:
[0023] Y=(NM) 2 (1)
[0024] 2.4) According to the above loss value Y, the connection weights between the residual layer and the identity mapping layer of the residual neural network, the threshold of the hidden layer and the threshold of the output layer are updated respectively;
[0025] 2.5) Repeat steps 2.2) to 2.4). When the loss value Y between the output value M of the residual neural network and the corresponding azimuth action N is less than the preset error threshold, the training of the residual neural network is completed to obtain the trained residual neural network.
[0026] In step 3), the test set is used to evaluate the trained residual neural network to determine whether the trained residual neural network can achieve the expected effect. If the evaluation result is yes, the optimized residual neural network is obtained. Otherwise, the specific method of returning to step 1) is as follows:
[0027] 3.1) Inputting the local environment conditions in the test set into the trained residual neural network to obtain the output data set X to be detected, wherein the specific data in the output data set X to be detected is the azimuth action output by the trained residual neural network;
[0028] 3.2) Using formula (2), calculate the mean square error W between the azimuth action in the output data set X to be tested and the corresponding azimuth action Z in the test set, and use the mean square error W to measure the overall error degree between the output of the residual neural network in the test set and the true output;
[0029]
[0030] Where n is the total number of data in the output data set X to be detected;
[0031] 3.3) When the mean square error W is less than 0.01, it indicates that the expected effect of the trained residual neural network has been achieved, and the optimized residual neural network is obtained, otherwise return to step 1).
[0032] In step 4), the sensors include laser sensors, ultrasonic sensors, lidar, depth cameras, and GPS.
[0033] In step 5), based on the above-mentioned planned azimuth action, the corresponding optimal angular velocity and optimal linear velocity are obtained by the dynamic window method, and then the unmanned vehicle moves to the next position according to the optimal angular velocity and optimal linear velocity. The specific method is as follows:
[0034] 5.1) Establish the kinematic model of the unmanned vehicle regarding speed and angular velocity as shown in formula (3):
[0035]
[0036] Among them, (x, y) is the current position coordinate of the unmanned vehicle, θ t is the heading angle of the unmanned vehicle, v is the linear velocity of the unmanned vehicle, w is the angular velocity of the unmanned vehicle, Δt is the movement time interval of the unmanned vehicle, which is 3 seconds;
[0037] 5.2) According to the physical constraints including the current speed and acceleration of the unmanned vehicle, determine the speed space that the unmanned vehicle can reach, and sample a pair of speed combinations (v, w) in the speed space;
[0038] 5.3) The current position coordinates (x, y) and heading angle θ of the unmanned vehicle t , and the above speed combination (v, w) is substituted into the kinematic model of the unmanned vehicle regarding speed and angular velocity to obtain the motion trajectory of the unmanned vehicle within the motion time interval Δt; a series of speed combinations (v, w) are obtained by sampling, corresponding to a series of motion trajectories of the unmanned vehicle;
[0039] 5.4) Using a comprehensive evaluation function to select the optimal motion trajectory from the series of unmanned vehicle motion trajectories obtained above;
[0040] 5.5) Obtaining the speed combination corresponding to the optimal motion trajectory can obtain the optimal angular velocity and optimal linear velocity;
[0041] 5.6) The unmanned vehicle moves to the next position according to the optimal angular velocity and optimal linear velocity.
[0042] The formula of the comprehensive evaluation function is:
[0043] G(v,w)=σ(α*heading(v,w)+β*dist(v,w)+γ*velocity(v,w)) (4)
[0044] Among them, heading(v,w) is the heading angle evaluation function, dist(v,w) is the obstacle distance evaluation function, velocity(v,w) is the speed evaluation function, and σ, α, β, and γ are weight coefficients. The sum of the three evaluation functions constitutes a comprehensive evaluation function, which is used as a standard to screen the optimal motion trajectory within the time window.
[0045] The unmanned vehicle local path planning method based on neural network and dynamic window method provided by the present invention has the following beneficial effects: using neural network to plan the next action of the unmanned vehicle avoids the problems of local optimum and narrow space being difficult to handle, and ensures the target reachability in complex environments; using dynamic window method to actually control the movement of the unmanned vehicle, and ensures the absolute safety of the unmanned vehicle in avoiding obstacles; having autonomous learning ability, and being able to adapt to various different working environments; and directly planning in speed space is also convenient for direct use in actual deployment. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] Figure 1 A flow chart of the local path planning method for an unmanned vehicle based on a neural network and a dynamic window method provided by the present invention.
[0047] Figure 2 The flowchart of the process of establishing a data set in the present invention.
[0048] Figure 3 Schematic diagram of the azimuth action of the current sampling point in the present invention.
[0049] Figure 4 It is a schematic diagram of the process of obtaining the optimal angular velocity and the optimal linear velocity by the dynamic window method in the present invention. DETAILED DESCRIPTION
[0050] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.
[0051] like Figure 1 As shown, the unmanned vehicle local path planning method based on neural network and dynamic window method provided by the present invention includes the following steps performed in sequence:
[0052] 1) Based on the working environment of the unmanned vehicle, the grid map is generated in batches and the A* algorithm is used to plan the global optimal path. Then, the global optimal path is sampled to obtain various local environmental conditions and corresponding azimuth actions, a data set is established, and the data set is divided into a training set and a test set in proportion;
[0053] The specific method is as follows:
[0054] 1.1) If Figure 2 As shown, the working environment of the unmanned vehicle is selected, a map of the working environment is constructed, and geographic information including terrain and objects is extracted from the map;
[0055] 1.2) Input the above geographic information into map software such as ArcGIS and GIS to batch generate a large number of raster maps with similar geographic information, then randomly specify the starting point and target point in the raster map, and use the A* algorithm to plan the global optimal path;
[0056] 1.3) In the above global optimal path, a sampling point is set at a certain distance as the current sampling point, and the map information matrix within the 7×7 grid range around the current sampling point is collected as the local environment of the current sampling point;
[0057] 1.4) For each grid in the above map information matrix, if there is an obstacle in the grid, -1 is marked in the grid, otherwise the Euclidean distance from the center point of the grid to the target point is marked in the grid;
[0058] 1.5) Calculate the angle between the line connecting the next sampling point and the current sampling point and the horizontal direction and use it as the azimuth of the current sampling point, such as Figure 3 As shown;
[0059] 1.6) Taking the local environment of the current sampling point as input and the corresponding azimuth action as output, a complete set of training data is formed. Sampling all the global optimal paths in the batch-generated raster maps can obtain sufficient training data. All the training data constitute a data set, and the data set is divided into a training set and a test set at a ratio of 9:1.
[0060] 2) constructing a residual neural network, and then iteratively optimizing the residual neural network using the training data in the above training set to obtain a trained residual neural network;
[0061] The specific method is as follows:
[0062] 2.1) Construct a residual neural network including an input layer, a pooling layer, a residual layer, an identity mapping layer and an output layer;
[0063] The training data of the training set in the present invention is large in scale, and a deep neural network is required to learn and memorize all the training data. When the number of neural network layers is deep, the performance of the residual neural network is better, so the present invention uses the residual neural network.
[0064] The number of residual layers and identity mapping layers in the residual neural network is 20-30 layers, and the residual layer and the identity mapping layer are fully connected. The performance of the residual neural network will get better as the number of layers increases. When the number of residual layers and identity mapping layers is 20, the residual neural network already has good performance for path planning, and the deeper the number of layers, the better the performance. However, when the number of residual layers and identity mapping layers reaches 30, the performance of the residual neural network will not be significantly improved.
[0065] 2.2) Input the local environment conditions in the training set into the above residual neural network to obtain the output value M of the residual neural network;
[0066] 2.3) The mean square error loss function shown in Formula 1 is used to calculate the output value M of the residual neural network and the loss value Y of the corresponding azimuth action N in the training set:
[0067] Y=(NM) 2 (1)
[0068] 2.4) According to the above loss value Y, the connection weights between the residual layer and the identity mapping layer of the residual neural network, the threshold of the hidden layer and the threshold of the output layer are updated respectively;
[0069] 2.5) Repeat steps 2.2) to 2.4), when the loss value Y between the output value M of the residual neural network and the corresponding azimuth action N is less than the preset error threshold, the training of the residual neural network is completed to obtain the trained residual neural network. The preset error threshold in the present invention is 0.01. A lower error threshold can obtain higher navigation accuracy. When the preset error threshold is lower than 0.01, the planning requirements can be met.
[0070] 3) Using the test set to evaluate the above trained residual neural network, whether the trained residual neural network can achieve the expected effect, if the evaluation result is yes, the optimized residual neural network is obtained, otherwise return to step 1);
[0071] The specific method is as follows:
[0072] 3.1) Inputting the local environment conditions in the test set into the trained residual neural network to obtain the output data set X to be detected, wherein the specific data in the output data set X to be detected is the azimuth action output by the trained residual neural network;
[0073] 3.2) Using formula (2), calculate the mean square error W between the azimuth action in the output data set X to be tested and the corresponding azimuth action Z in the test set, and use the mean square error W to measure the overall error degree between the output of the residual neural network in the test set and the true output;
[0074]
[0075] Where n is the total number of data in the output data set X to be detected;
[0076] 3.3) When the mean square error W is less than 0.01, it indicates that the expected effect of the trained residual neural network has been achieved, and the optimized residual neural network is obtained, otherwise return to step 1).
[0077] 4) Using the sensors on the unmanned vehicle to scan the current environment information to obtain the map information matrix, and then input the above-mentioned optimized residual neural network for path planning to obtain the corresponding planned azimuth action;
[0078] Unmanned vehicles are usually equipped with various types of sensors, such as laser sensors, ultrasonic sensors, laser radars, depth cameras, GPS, etc. In the present invention, laser sensors and GPS are set on the unmanned vehicle to establish the map information matrix of the unmanned vehicle. The laser sensor can detect the position of obstacles around the unmanned vehicle and determine the obstacle grid in the map information matrix; the GPS can determine the current coordinates of the unmanned vehicle, and calculate the distance between each grid and the target point by calculating with the coordinates of the target point, so as to determine the distance grid in the map information matrix. The map information matrix obtained above is input into the optimized residual neural network, and the planned azimuth action θ output by the optimized neural network can be obtained.
[0079] 5) Based on the above-mentioned planned azimuth action, the corresponding optimal angular velocity and optimal linear velocity are obtained by a dynamic window method, and then the unmanned vehicle moves to the next position according to the optimal angular velocity and optimal linear velocity;
[0080] The specific method is as follows:
[0081] 5.1) Establish the kinematic model of the unmanned vehicle regarding speed and angular velocity as shown in formula (3):
[0082]
[0083] Among them, (x, y) is the current position coordinate of the unmanned vehicle, θ t is the heading angle of the unmanned vehicle, v is the linear velocity of the unmanned vehicle, w is the angular velocity of the unmanned vehicle, Δt is the movement time interval of the unmanned vehicle, which is 3 seconds;
[0084] 5.2) According to the physical constraints including the current speed and acceleration of the unmanned vehicle, determine the speed space that the unmanned vehicle can reach, and sample a pair of speed combinations (v, w) in the speed space;
[0085] 5.3) The current position coordinates (x, y) and heading angle θ of the unmanned vehicle t , and the above speed combination (v, w) is substituted into the kinematic model of the unmanned vehicle regarding speed and angular velocity to obtain the motion trajectory of the unmanned vehicle within the motion time interval Δt; a series of speed combinations (v, w) are obtained by sampling, corresponding to a series of motion trajectories of the unmanned vehicle;
[0086] 5.4) Using a comprehensive evaluation function to select the optimal motion trajectory from the series of unmanned vehicle motion trajectories obtained above;
[0087] 5.5) Obtaining the speed combination corresponding to the optimal motion trajectory can obtain the optimal angular velocity and optimal linear velocity;
[0088] 5.6) The unmanned vehicle moves to the next position according to the optimal angular velocity and optimal linear velocity.
[0089] The formula of the comprehensive evaluation function is:
[0090] G(v,w)=σ(α*heading(v,w)+β*dist(v,w)+γ*velocity(v,w)) (4)
[0091] Among them, heading(v,w) is the heading angle evaluation function, dist(v,w) is the obstacle distance evaluation function, velocity(v,w) is the speed evaluation function, and σ, α, β, and γ are weight coefficients. The sum of the three evaluation functions constitutes a comprehensive evaluation function, which is used as a standard to screen the optimal motion trajectory within the time window.
[0092] The heading angle evaluation function is specifically as follows: assuming that the heading angle at the end of the time window trajectory under the current speed combination is φ, the smaller the difference between this direction and the planned azimuth action θ, the higher the heading angle evaluation function value, indicating that the motion trajectory corresponding to this speed combination is better.
[0093] The obstacle distance evaluation function is specifically as follows: let η be the minimum distance between the trajectory curve corresponding to the current speed combination and the obstacle vehicle. The larger the value, the higher the obstacle distance evaluation function value, indicating that the motion trajectory corresponding to this speed combination is better.
[0094] The specific speed evaluation function is: the closer the unmanned vehicle is to the target speed during normal driving, the better. Therefore, the higher the speed, the higher the value of the speed evaluation function.
[0095] 6) When the unmanned vehicle is moving, it continuously detects whether the target point appears in the detection window and there are no obstacles between the unmanned vehicle and the target point; if the detection result is yes, the navigation is considered to be completed, and the unmanned vehicle moves directly to the target point in a straight line and ends the navigation; otherwise, after the unmanned vehicle moves for a certain period of time at the optimal angular velocity and optimal linear velocity corresponding to the optimal trajectory, steps 4)-5) are repeated, and the next planning process is immediately started, and the vehicle moves along the optimal motion trajectory planned next time, and this process is repeated until the navigation is completed.
Claims
1. A local path planning method for an unmanned vehicle based on a neural network and a dynamic window method, characterized in that: The unmanned vehicle local path planning method based on neural network and dynamic window method comprises the following steps performed in sequence: 1) Based on the working environment of the unmanned vehicle, the grid map is generated in batches and the A* algorithm is used to plan the global optimal path. Then, the global optimal path is sampled to obtain various local environmental conditions and corresponding azimuth actions, a data set is established, and the data set is divided into a training set and a test set in proportion; 2) constructing a residual neural network, and then iteratively optimizing the residual neural network using the training data in the above training set to obtain a trained residual neural network; 3) Using the test set to evaluate the above trained residual neural network, whether the trained residual neural network can achieve the expected effect, if the evaluation result is yes, the optimized residual neural network is obtained, otherwise return to step 1); 4) Using the sensors on the unmanned vehicle to scan the current environment information to obtain the map information matrix, and then input the above-mentioned optimized residual neural network for path planning to obtain the corresponding planned azimuth action; 5) Based on the above-mentioned planned azimuth action, the corresponding optimal angular velocity and optimal linear velocity are obtained by a dynamic window method, and then the unmanned vehicle moves to the next position according to the optimal angular velocity and optimal linear velocity; 6) When the unmanned vehicle is moving, it continuously detects whether the target point appears in the detection window and there are no obstacles between the unmanned vehicle and the target point; if the detection result is yes, the navigation is considered to be completed, and the unmanned vehicle moves directly to the target point in a straight line and ends the navigation; otherwise, after the unmanned vehicle moves for a certain period of time at the optimal angular velocity and optimal linear velocity corresponding to the optimal trajectory, steps 4)-5) are repeated, and the next planning process is immediately started, and the vehicle moves along the optimal motion trajectory planned next time, and this process is repeated until the navigation is completed.
2. The unmanned vehicle local path planning method based on neural network and dynamic window method according to claim 1 is characterized in that: In step 1), based on the working environment of the unmanned vehicle, the grid map is generated in batches and the global optimal path is planned using the A* algorithm, and then the global optimal path is sampled to obtain various local environmental conditions and corresponding azimuth actions, a data set is established, and the specific method of dividing the data set into a training set and a test set in proportion is as follows: 1.1) Select the working environment of the unmanned vehicle, build a map of the working environment, and extract geographical information including terrain and objects from the map; 1.2) Input the above geographic information into map software such as ArcGIS and GIS to batch generate a large number of raster maps with similar geographic information, then randomly specify the starting point and target point in the raster map, and use the A* algorithm to plan the global optimal path; 1.3) In the above global optimal path, a sampling point is set at a certain distance as the current sampling point, and the map information matrix within the 7×7 grid range around the current sampling point is collected as the local environment of the current sampling point; 1.4) For each grid in the above map information matrix, if there is an obstacle in the grid, -1 is marked in the grid, otherwise the Euclidean distance from the center point of the grid to the target point is marked in the grid; 1.5) Calculate the angle between the line connecting the next sampling point and the current sampling point and the horizontal direction and use it as the azimuth angle of the current sampling point; 1.6) Taking the local environment of the current sampling point as input and the corresponding azimuth action as output, a complete set of training data is formed. Sampling all the global optimal paths in the batch-generated raster maps can obtain sufficient training data. All the training data constitute a data set, and the data set is divided into a training set and a test set at a ratio of 9:
1.
3. The unmanned vehicle local path planning method based on neural network and dynamic window method according to claim 1, characterized in that: In step 2), the residual neural network is constructed, and then the residual neural network is iteratively optimized using the training data in the training set to obtain the trained residual neural network. The specific method is as follows: 2.1) constructing a residual neural network including an input layer, a pooling layer, a residual layer, an identity mapping layer and an output layer; the number of residual layers and identity mapping layers in the residual neural network is 20-30 layers, and the residual layer and the identity mapping layer are fully connected; 2.2) Input the local environment conditions in the training set into the above residual neural network to obtain the output value M of the residual neural network; 2.3) The mean square error loss function shown in Formula 1 is used to calculate the output value M of the residual neural network and the loss value Y of the corresponding azimuth action N in the training set: Y=(N-M) 2 (1) 2.4) According to the above loss value Y, the connection weights between the residual layer and the identity mapping layer of the residual neural network, the threshold of the hidden layer and the threshold of the output layer are updated respectively; 2.5) Repeat steps 2.2) to 2.4). When the loss value Y between the output value M of the residual neural network and the corresponding azimuth action N is less than the preset error threshold, the training of the residual neural network is completed to obtain the trained residual neural network.
4. The unmanned vehicle local path planning method based on neural network and dynamic window method according to claim 1, characterized in that: In step 3), the test set is used to evaluate the trained residual neural network to determine whether the trained residual neural network can achieve the expected effect. If the evaluation result is yes, the optimized residual neural network is obtained. Otherwise, the specific method of returning to step 1) is as follows: 3.1) Inputting the local environment conditions in the test set into the trained residual neural network to obtain the output data set X to be detected, wherein the specific data in the output data set X to be detected is the azimuth action output by the trained residual neural network; 3.2) Using formula (2), calculate the mean square error W between the azimuth action in the output data set X to be tested and the corresponding azimuth action Z in the test set, and use the mean square error W to measure the overall error degree between the output of the residual neural network in the test set and the true output; Where n is the total number of data in the output data set X to be detected; 3.3) When the mean square error W is less than 0.01, it indicates that the expected effect of the trained residual neural network has been achieved, and the optimized residual neural network is obtained, otherwise return to step 1).
5. The unmanned vehicle local path planning method based on neural network and dynamic window method according to claim 1, characterized in that: In step 4), the sensors include laser sensors, ultrasonic sensors, lidar, depth cameras, and GPS.
6. The local path planning method for an unmanned vehicle based on a neural network and a dynamic window method according to claim 1, characterized in that: In step 5), based on the above-mentioned planned azimuth action, the corresponding optimal angular velocity and optimal linear velocity are obtained by the dynamic window method, and then the unmanned vehicle moves to the next position according to the optimal angular velocity and optimal linear velocity. The specific method is as follows: 5.1) Establish the kinematic model of the unmanned vehicle regarding speed and angular velocity as shown in formula (3): Among them, (x, y) is the current position coordinate of the unmanned vehicle, θ t is the heading angle of the unmanned vehicle, v is the linear velocity of the unmanned vehicle, w is the angular velocity of the unmanned vehicle, Δt is the movement time interval of the unmanned vehicle, which is 3 seconds; 5.2) According to the physical constraints including the current speed and acceleration of the unmanned vehicle, determine the speed space that the unmanned vehicle can reach, and sample a pair of speed combinations (v, w) in the speed space; 5.3) The current position coordinates (x, y) and heading angle θ of the unmanned vehicle t , and the above speed combination (v, w) is substituted into the kinematic model of the unmanned vehicle regarding speed and angular velocity to obtain the motion trajectory of the unmanned vehicle within the motion time interval Δt; a series of speed combinations (v, w) are obtained by sampling, corresponding to a series of motion trajectories of the unmanned vehicle; 5.4) Using the comprehensive evaluation function to select the optimal motion trajectory of the series of unmanned vehicle motion trajectories obtained above; 5.5) Obtaining the speed combination corresponding to the optimal motion trajectory can obtain the optimal angular velocity and optimal linear velocity; 5.6) The unmanned vehicle moves to the next position according to the optimal angular velocity and optimal linear velocity.
7. The unmanned vehicle local path planning method based on neural network and dynamic window method according to claim 6 is characterized in that: The formula of the comprehensive evaluation function is: G(v,w)=σ(α*heading(v,w)+β*dist(v,w)+γ*velocity(v,w)) (4) Among them, heading(v,w) is the heading angle evaluation function, dist(v,w) is the obstacle distance evaluation function, velocity(v,w) is the speed evaluation function, and σ, α, β, and γ are weight coefficients. The sum of the three evaluation functions constitutes a comprehensive evaluation function, which is used as a standard to screen the optimal motion trajectory within the time window.
Citation Information
Cited By
Unmanned aerial vehicle optimal path planning method and system combined with neural network
CN122281928A
An unmanned aerial vehicle optimal path planning method and system combined with a neural network
CN122281928B
