Research On Key Technologies Of Autonomous Driving Based On Deep Learning And Deep Reinforcement Learning

In recent years,with the continuous development of artificial intelligence,autonomous driving has also made great progress as its important landing project.At present,autonomous driving is mainly composed of three aspects: environmental perception,behavior decision and vehicle communication.This paper proposes a real-time semantic segmentation algorithm based on deep learning for visual assistance in environmental perception,and performs real-time semantic analysis on street view images in the driving direction;for behavioral decisions in autonomous driving,a deep reinforcement learning method is proposed to optimize;In terms of improving the communication efficiency between the vehicle and the roadside unit in vehicle communication,a method based on deep reinforcement learning is proposed to jointly optimize the intelligent active buffering of the roadside unit in the case of vehicle automatic driving.The main work and innovations of this thesis are as follows:1.Aiming at the problems of large amount of computation and insufficient feature utilization when using a convolutional neural network to semantically segment street view images,a coding-decoding semantic segmentation network structure based on lightweight multi-branch feature cross-layer fusion is proposed.The structure uses original images of different resolutions as the input of multi-branch in the network,and after input,uses spatial multi-purpose convolution kernel to initially extract features,and then uses the residual module to perform subsequent feature extraction and feature fusion,and Use deep separable convolution instead of traditional convolution to reduce the amount of network computation.Between each branch of the network,feature fusion at each stage is carried out,and the features extracted after the branch ends are transferred to the next branch.It is verified by experiments that the method achieves a processing speed of 112.3FPs in the Cityscapes data set,and can guarantee 65.6 % mIoU semantic segmentation performance.It meets the requirements of real-time semantic segmentation.2.Aiming at the behavior decision problem of autonomous driving,the Markov decision process is used for modeling,and an automatic driving behavior decision algorithm based on deep reinforcement learning is proposed.First,the system model is set to a one-way straight three-lane highway that contains multiple interfering vehicles.The driving behavior is subdivided into: lane change to the left,lane change to the right,speed changed to 0(stop),speed changed to 1,speed changed to 2,speed changed to 3.Take the state around the vehicle as the result of its environmental perception and set it to be obtained from the system model.Then set the corresponding reward value for different states and actions.An automatic driving behavior decision algorithm based on deep reinforcement learning is proposed.Finally,the experiment proves that the method achieves good performance under different parameters such as learning rate and number of interfering vehicles.3.Aiming at the problem of improving the communication efficiency between the vehicle and the roadside unit,an optimization scheme for actively caching the roadside unit while the vehicle is driving automatically is proposed.At this time,two independent Markov decision processes are used to model the two objects,vehicle and roadside unit,respectively.On the basis of the previous method,the system model has placed multiple roadside units on the side of the road,and it can evenly cover the driving area of the vehicle.The vehicle receives data from the roadside unit according to the received packet rate calculated by the corresponding wireless channel model.The roadside unit determines the amount of cached data based on the vehicle’s traveling speed,the received data,and the cache of the previous roadside unit.A dual DQN joint deep reinforcement learning algorithm is proposed to optimize two different decision-making processes.After experimental verification,the method achieves good performance under different parameters such as learning rate,data packet demand,roadside unit distance and other parameters.Finally,the shortcomings of the above methods and the direction of improvement in the future are analyzed.