Research On Several Methods Of Unmanned Driving And Decision Based On Deep Learning

Driverless is a very hot technology,mainly through computer systems and some smart sensors to achieve the purpose of driverless.It includes three aspects of perception,decision and control.Perceptual aspects such as lane recognition and vehicle identification still have many problems,such as real-time problems,illumination occlusion problems,point cloud data processing problems,two-dimensional vehicle identification and lane line recognition are not specific enough to describe the driving environment.The decision-making aspect has strict requirements for safety and reliability.At present,most of the decision-making algorithms are constructed according to the rules,only to deal with the normal driving environment,and it is difficult to make timely and correct decisions in the face of various emergencies in the driving environment.In this paper,the related research on the driverless perception and decision-making part is carried out.The main contents of the research are as follows:(1)The traditional lane recognition algorithm is mostly based on the lane detection of traditional image processing technology.However,the traditional algorithm needs to perform preprocessing,edge detection,Hough transform and other steps.Each step is independent but affects each other.The overall optimal purpose,and the real-time problem also makes the application of the algorithm not high.The most important thing is that the traditional lane recognition method is not suitable for dealing with lane changes and complex driving environment.In this paper,a lane recognition method based on full convolutional neural network and conditional random field is proposed for the above problems.The method transforms the lane recognition task based on the lane line into an image segmentation task,and the segmentation process of the entire lane reaches an end-to-end structure,and has self-learning ability for input various driving environment pictures to adapt to lane detection in various complicated environments.At the same time,the lane detection technology based on image segmentation is more in line with the real driving environment,and has a good separation effect for obstacles such as various pedestrian vehicles in the lane.(2)Most of the traditional vehicle identification algorithms are based on 2D image detection,although some related network algorithms such as Faster RCNN,YOLO,SSD,etc.in deep learning have achieved high accuracy in 2D images.And for the analysis of unmanned scenes also have a good effect.However,the description of the3 D real-world scene is still not enough.For example,in the unmanned driving,inaddition to detecting pedestrians,vehicles,and obstacles,the detection and positioning of the speed and direction of the object are also very important.To solve these problems,this paper uses the point cloud data of optical radar combined with RGB image to realize the detection task of 3D vehicle to describe the real driving scene more completely,and utilizes PointNet network to deal with the processing technology of point cloud data and high computational complexity.In the detection process,thesis uses the CNN network and the cascading rejection classifier and the correlation real-time and accuracy advantages of the column correlation to improve the vehicle detection effect in the 3D environment.(3)Most of the existing unmanned decision-making algorithms adopt the construction of rules,and specific rules are defined for specific driving scenarios.However,in the real scene,it is difficult for traditional algorithms to make correct decisions completely due to complex and variegated environments such as climate traffic and difficult situations.In order to solve these problems and make the driverless system more intelligent decision-making scheme,thesis proposes a decision-making algorithm based on reinforcement learning.Reinforcement learning is an unsupervised machine learning algorithm.Compared with most deep learning algorithms,it does not need Manual labeling eliminates a lot of cost.In the thesis,the RGB image input is used as the state space,and the steering brake of the vehicle is defined as the behavior space.At the same time,the distance is completed as a reward and punishment mechanism without violating the traffic rules and safety,so that the algorithm model can spontaneously travel.Learning decision-making methods in driving.