Behavior Planning Method Of Unmanned Vehicle Based On Autonomous Learning

Motion planning is the key technology of autonomous driving for unmanned vehicle.Vision based motion planning provides an inexpensive solution.However,the complexity of road image distribution and the huge amount of image compution make it difficult to design a motion planning system of unmanned vehicle.In this paper,the design of robust and real-time motion planning system for unmanned vehicle is studied.Aiming at the technical difficulties in the design process,a solution is proposed to solve the difficulties of modeling and computing directly in the image space.It can plan the motion of unmanned vehicle better and prevent the occurrence of error control effectively.The main work of this article is as follows:Firstly,a motion planning system for unmanned vehicle based on heterogeneous deep learning is proposed.This system designs different kinds of deep learning models.The dimension of road image is descended firstly and the image space is converted to standard normal distribution space.Then the motion planning for unmanned vehicle is realized in the coding space.This system includes an auto encoding model based on convolutional neural networks excelling in image modeling,a road tracking model based on recurrent neural networks excelling in sequential data modeling,a control model and an evaluation model both based on convolutional neural networks,SENet or ResNet.Secondly,based on the motion planning system with heterogeneous deep learning,a motion planning method for unmanned vehicle based on finite state machine and deep learning is proposed.Local path planning is completed by the motion planning system with heterogeneous deep learning.The motion selection for a given global path or an intersection needs to be completed with a finite state machine.Through the analysis of global path,a finite state machine is used to complete the motion selection in road tracking or an intersection.With this system,the difficulty of training a single control model is reduced and the global path can be completed based on the finite state machine.Finally,based on the motion planning system with heterogeneous deep learning,a motion planning method for unmanned vehicle based on collaborative deep learning is proposed.In the motion planning system with heterogeneous deep learning,only color images are used as data source.In this system,control model could learn the ability of motion planning from two views of dataset.It can use a large number of unlabeled samples to improve the utilization of samples and the generalization ability of the whole model.The experimental results show that the evaluation model can evaluate the control model effectively,and prevent the system from making wrong decisions.In addition,this system not only learns the abilities provided by training set,such as road tracking and turning,but also learns the ability of local obstacle avoidance and lane correction with a certain generalization ability.