Path Replanning Of Unmanned Surface Vehicle Based On Deep Learning

Effective and reasonable development and detection of marine resources is a challenge with opportunity.In the complex marine environment,the applicability of unmanned surface vehicle,different from UAV and mobile robots,has been further highlighted,and its development has attracted more and more attention from scholars at home and abroad.The path planning of USV has also become an indispensable key in the development process.The purpose is to plan the optimal path through the optimization algorithms.A large number of planning algorithms have been accumulated in this field,including A*algorithm,Fast marching method,Artificial potential field method,Genetic algorithm,Ant colony algorithm,and Fast extended search Random Tree algorithm.The above algorithms have problems such as path optimization and poor dynamic generation performance caused by the lack of information.At the same time,in the analysis and research of the maritime collision avoidance rules,it is still in the unformed stage in the navigation avoidance of USV.With the great improvement of computer computing efficiency,the development and application of neural networks have been further improved.The idea of neural network and deep learning is more and more used for path planning and control algorithms of USV,and has superior engineering applicability.This paper emphazes the optimization of the USV path planning algorithm,summarizes the traditional path planning algorithm of USV,and focues on the deep learning algorithms to develop the CNN-LSTM based planning algorithm,taking full advantage of the CNN and LSTM neural networks to excavate the image space and time series ability,respectively.This paper constructs a CNN-LSTM(C-LSTM)path planning algorithm framework.It uses hierarchical idea to handle environmental information,separating dynamic environmental information from static field,and combines its speed,heading,and other information to vectorize obstacle ship areas.The object model considers the COLLEGS,improves the USV obstacle-avoidance ability and the decision-making ability combined with maritime rules.In this paper,based on the structure of Fast Marching method,combined with the information of obstacle ship speed and heading angle,an improved Constrained C-FM algorithm is proposed and verified by simulation test.Initially,there is the effectiveness of the path replanning algorithm.Further,this paper establishes a deep learning model including Convolutional neural networks and Long Short Term Memory neural network.The formula deduces that LSTM effectively solves the gradient disappearance and explosion phenomenon of RNN algorithm,and has strong theoretical value.On this basis,the algorithm parameter configuration suitable for the path planning problem is proposed.Under the support of the literature data,the algorithm is validated and the accuracy is obtained.This paper constructs a path planning algorithm framework based on C-LSTM,and decipts a training and verification data suitable for complex marine environments.In order to further improve the accuracy of the network structure,this paper discusses and analyzes the parameter configuration problems such as algorithm over-fitting,hidden layer unit number and activation function selection to ensure the effectiveness of the algorithm.Research and analysis of C-LSTM-based static path generation capability are carried out.The high quality path generated by C-LSTM algorithm in global path planning is better than FM algorithm and LSTM algorithm,which further validates the feasibility of the algorithm.At the same time,this paper concentrates on the simulation experiments of dynamic path generation based on the C-LSTM algorithm.This paper conducts six kinds of dynamic navigation simulation tests,including static obstacles and dynamic navigation ships,which simulate the known and unknown environment,head on,overtake and cross situations,the obstacle-avoidance characteristics of the dynamic navigation of proposed algorithm are analyzed,and its path generation ability will be verified.The results show that,in the known environment,the LSTM algorithm has better speed and heading deviation than the improved constrained C-FM algorithm,and introduces the CNN algorithm to achieve an unknown environment(irregular obstacles)excellent navigation stability.At the same time,the C-LSTM algorithm shows good path re-planning ability under the all conditions and can better adjust the speed heading information and effectively follow the COLREGS.In addition,in the cross stuation,after the obstacle ship does not implement the avoidance strategy,the C-LSTM algorithm uses the speed reduction strategy to complete the avoidance of the obstacle Vehicle.