Research On Cooperative Perception Techniques In C-V2X Scenario

Cooperative perception technology can be applied in the Vehicle-to-Vehicle network to share perception data from sensors such as in-vehicle cameras and lidars among multiple vehicles,thereby extending the limited perception range of individual vehicles.In theory,frequent exchange of surrounding environment perception data among vehicles can effectively improve the accuracy and reliability of vehicle perception.However,channel and computational resources are often limited in Vehicle-to-Vehicle network,making it a key challenge to design an effective multi-vehicle cooperative perception solution under resource constraints.As cellular base stations have larger bandwidth,stronger transmit and receive power,and more antennas,the current V2V(Vehicle to Vehicle)communication architecture is gradually evolving towards the C-V2V(Cellular Vehicle to Vehicle)architecture based on cellular network communication.In the cellular vehicle network scenario,this article studies the two key issues of perception compression and perception sharing in cooperative perception among vehicles,with the main research content as follows:Firstly,this thesis proposes a domain-adaptive deep image compression method for the perception compression problem in cooperative perception.The method effectively achieves the transfer of a general deep image compression network to a target domain deep image compression network,based on the Gaussian mixture domain adaptation parameters and fine-grained bottleneck scale quantization parameters.Under the premise of freezing the autoencoder and hyperprior network parameters,adding only an additional Gaussian mixture parameter layer and fine-grained bottleneck scale parameter layer can achieve domain adaptation for image compression from general image compression to a specific domain and realize discrete rates of different compression levels.Secondly,this thesis proposes a cooperative perception decision-making method based on deep reinforcement learning to address the perception sharing problem.In the C-V2 V scenario,an edge server connected to micro base stations is introduced to achieve centralized scheduling of cooperative perception information.The original combined optimization problem of cooperative perception is decoupled into a Markov decision process,and a solution based on deep reinforcement learning is designed.This approach effectively improves the perception synergy and has the ability to generalize to different traffic scenarios.Based on the aforementioned research,this thesis designs and implements a cooperative perception simulation framework based on the cellular vehicle network,including a deep image compression system for traffic scenes and a cooperative perception decision-making system based on simulated point clouds,and then we demonstrates them and verifies their effectiveness.