Research On Perception And Communication For Intelligent Connected Vehicles Based On Deep Learning

The development of Vehicular Network(VN)and Artificial Intelligence(AI)technology has promoted the connection and intelligence of unmanned vehicles.Intelligent Connected Vehicles(ICVs)are an important part of the future Smart City system.Intelligent connected vehicles rely on their own sensors to perceive the surrounding environment and interact with other ICVs through wireless communication technology.Finally,ICVs combine perception results with communication information to make complex decisions and plans.However,the perception ability and the quality of services(Qo S)of communication of ICVs,due to the increase of transport participants and the influence of weather and road conditions,are facing the following problems: communication bandwidth and resource limitations of vehicular edge computing(VEC),hindered sensing capabilities for long distances,complex vehicle networking protocol restrictions,etc.This study uses the advantages of deep learning(DL)and reinforcement learning(RL)methods to construct a method that meets the perception accuracy and communication Qo S of ICVs to solve these challenges.The main contributions are listed as follows:(1)To overcome the limitation of local surrounding perception,this study proposes a collaborative perception framework to research the potential of the combination of3 D target detection and communication technology and gives a cooperative perception method for ICVs based on Pillar.After the ICVs encodes the point cloud data based on Pillar,this framework uses Octomap to compress and transmit sensing feature data.ICVs can effectively reduce blind spots in environmental perception and improve the accuracy of long-distance 3D target detection by this cooperative approach.Experimental results show that the proposed method outperforms other known cooperative perception schemes in terms of processing frequency and computing time,and the accuracy of object detection is significantly improved at further distances.In addition,this method realizes the overall low latency of fusion data processing and cooperative perception message transmission.This method provides more sufficient decision-making reaction time for ICVs under the premise of meeting the accuracy requirements of perception tasks.In addition,this method also provides a unified information interaction scheme for the cooperative perception of different ICVs.(2)This study proves a perception task offloading method for VEC based on constrained hybrid action space.Considering the constraints of perception tasks for ICVs and the influence of the mixed action space,this study proposes a deep reinforcement learning(DRL)method which can handle the mixed action space with constraints.This method is based on the classic reinforcement learning algorithm-Proximal Policy Optimization(PPO).VEC network can slice,and offload sensing tasks based on the different states of edge computing servers.Besides,the proposed method can meet the constraints of sensing task segmentation.In addition,it also solves the multi-objective optimization problem in three aspects: the offloading of the perception task of the network,the calculation consumption of the perception task,and the energy consumption of the transmission tasks.The experimental results show that the proposed method can enable ICVs to offload tasks with low energy consumption and low latency.Compared with existing methods,the proposed method has greatly improved and improved the perception application service quality of ICVs(3)A MAC layer method for Intelligent Connected Vehicles based on sensinginformation-value.Aiming at the different delay requirements of sensing tasks,a reciprocal partially observable MAC access method based on the value of sensing information is proposed.In this study,a reciprocal cooperative reward function is proposed by considering the value of different sensing tasks and waiting time for transmission,that is,the current transmission probability of ICVs not only considers its own revenue value,but also considers the revenue of other connected vehicles in the entire VN.In this paper,a cross-layer cooperative game MAC layer model of ICVs is constructed.Furthermore,the existence of a game equilibrium point is theoretically proved using a partially observable Markov decision process(POMDP),and a specific method for controlling access control probabilities of ICVs using deep reinforcement learning(DRL)is provided.Finally,the experimental results of different sensing data arrival periods are given to evaluate the performance of the proposed access mechanism and compare it with the traditional IEEE 802.11 p standard protocol.The experimental results prove that the reciprocal benefit game method proposed in this study has the advantages of the delay in transmitting sensing data and the effective delivery rate.This method can strongly support the time-sensitive intelligent networked vehicle environment sensing tasks.