Research On Path Planning Based On Neural Network Selection

In the research on the application of unmanned vehicles,how to efficiently and safely plan the path for unmanned vehicles has always been a hot and challenging research content.The traditional path planning algorithms for unmanned vehicles mainly include global path planning algorithm and local path planning algorithm.The global path planning algorithm plans the obstacle avoidance navigation path for the unmanned vehicle by pre-processing the static obstacle information in the scene;The local path planning algorithm reads the dynamic obstacle information in the map scene through the laser radar sensor or infrared sensor of the unmanned vehicle,and provides obstacle avoidance path planning for the unmanned vehicle in real time.With the development of science and technology in the field of artificial intelligence,people train unmanned vehicles through deep reinforcement learning,and let unmanned vehicles plan obstacle avoidance navigation paths based on autonomous learning.The traditional algorithm can help the unmanned vehicle to complete the obstacle avoidance navigation task in a short time,but it is difficult to ensure the optimal path efficiency in the changeable scene.The deep reinforcement learning model can make the unmanned vehicle adapt to the changing scene,but it will encounter problems such as generating useless nodes in the planning process,and the training time is too long,resulting in low planning efficiency.Based on the in-depth analysis of the existing methods,this dissertation proposes a method of using neural network to train unmanned vehicles and dynamically select the best combination of efficiency and security from a variety of path planning algorithms.The main contributions of this dissertation are as follows:(1)A method of using neural network tools to train the combination of unmanned vehicle selection algorithms is proposed.A variety of global and local path planning algorithms provided by traditional algorithms can play a good role in a single scene through a fixed combination.The neural network method is used to select the optimal algorithm combination,which avoids the problem that the traditional method can not adapt to different map scenes because of the fixed combination.The proposed method is essentially a classification problem,and will not produce useless node branches and the "trial and error" process in reinforcement learning.(2)The simulation experiment of optimal path combination selection based on neural network is designed,and the optimization objective function PV is proposed.Using ros/gazebo simulation tool,2000 maps with random number and distribution of obstacles are constructed,and the car is allowed to try all algorithm combinations and carry out simulation experiments in all maps.The PV function is proposed to calculate the simulation data and select the optimal path planning algorithm combination.(3)Four kinds of neural network models are used to select the optimal algorithm combination.It is found that the model whose obstacle map information is input in the form of text has a good training effect.On this basis,the network is improved,the attention mechanism is introduced,and through the ablation experiment,it is found that the data can get the relatively optimal training effect through the attention module before entering the LSTM model.