Research On Multi-sensor Fusion Deep Reinforcement Learning For Autonomous Driving Tasks

Since the birth of automobiles,it has brought great convenience to human production and life while has also brought many negative effects,such as safety issues.The improvement of safety is one of the reasons that autonomous driving research has received so much attention in recent years.In addition,autonomous driving can improve efficiency and realize automated processes such as logistics,factories,and commodity transportation.In real world,autonomous vehicles are built by large teams of large companies through extensive engineering efforts.Nowadays,the end-to-end realization of the task of developing autonomous driving in deep learning can be realized by some intelligent algorithms.Based on this background,this paper combines representation learning and reinforcement learning to learn driving policy end-to-end in the CARLA autonomous driving simulator.The method of representation learning is used to process the data of different type of sensor and extract the system state,and the reinforcement learning algorithm take this state as input to output the control policy.The main work and contribution of this paper is as follows:(1)A multi-sensor fusion representation learning module is proposed in order to process different types of sensor.Specifically,an RGB image feature extraction network based on Vinsion Transformer,a depth map feature extraction network based on Dense Net and a point cloud data feature extraction network based on Point Net++are implemented.Those combined modules improves the efficiency of multi-source information fusion.(2)A reinforcement learning module for continuous action space in autonomous driving tasks is proposed.We also explored the effects of deterministic strategies and random strategies on different tasks,and designed effective reward functions for different tasks.(3)A training framework combining pre-training and offline training is proposed.The representation learning module is pre-trained on open source data sets CIFAR,Image Net,and Model Net,which improves the extraction of state information and reduces the instability of traditional reinforcement learning.The feature extraction of multisensor fusion improves the generalization of the system which makes this framework can be applied to richer scenarios.(4)A training framework for the CARLA simulator is built.We also conducted training on different autonomous driving tasks,and tested and verified the generalization ability of the proposed method among different scene maps.