Cooperative Sensing and Computation for Environment Perception in Autonomous Driving with Vehicular Edge Computing

Abstract

The advances in sensing technologies, artificial intelligence, Internet of vehicles, and edge computing paradigm pave the way for autonomous driving, which is a key use case that will re shape the future transportation systems in the 5G and 6G eras. Environment perception is a key module in autonomous driving that enables the autonomous vehicles (AVs) to view the surround ing environment, facilitating situational-aware decision and planning for autonomous driving. In this work, we consider a perception task for object detection and classification, and investigate a raw data level cooperative perception scheme, with cooperative sensing among AVs and cooperative computation among both edge server and AVs, to satisfy the stringent accuracy and delay requirements for the perception task with communication and computing resource efficiency. To exploit the differentiated sensing data quality at each AV for different objects, we partition the perception task into parallel object classification subtasks, and propose a differentiated data selection strategy which selects sensing data from different AVs for each subtask with accuracy satisfaction and resource efficiency. The computation of different subtasks is distributed in a vehicular edge computing network, in a communication efficient manner. An optimization problem is formulated for joint data selection, subtask placement and resource allocation, to minimize the total communication and computing resource consumption cost, while satisfying the delay and accuracy requirements. To facilitate data selection decision with accuracy satisfaction, a deep neural network (DNN) model is pre-trained to profile an accuracy estimation function, which estimates the accuracy for each object classification subtask given the data selection decision and the sensing data quality characterized by the sensing data volume and spatial diversity. An iterative solution based on genetic algorithm is proposed for the optimization problem. Simulation results demonstrate the accuracy improvement by the cooperative sensing strategy and the resource efficiency of the proposed differentiated data selection and subtask placement scheme.