Perception synergy optimization with deep reinforcement learning for cooperative perception in C-V2V scenarios

Abstract

Autonomous vehicles rely on sensors such as cameras and Lidar to collect data to make safe driving decisions. However, the sensors of a single-vehicle may be blocked or interfered with, and then the insufficient perception limits the safety of autonomous driving. Cooperative perception is to expand the perception domain of a single vehicle by sharing the perception data. Cooperative perception based on V2V (Vehicle to Vehicle) broadcast communication mode still might transmit a large amount of redundant data. Based on C-V2V (Cellular Vehicle to Vehicle) mmWave communication, fortunately, the edge server interconnected with the 5G small cell base station provides the possibility of centralized processing to leverage the perception interactive features among vehicles to carry on effective cooperative perception scheming. In this paper, a method based on deep reinforcement learning is designed to make centralized decisions to optimize the synergy of cooperative perception. In a highway scene, the road is firstly partitioned into several regions. As for each region, the interactive perception features of local vehicles and the regional perception features are obtained by an embedding method. Subsequently, according to the embedded features, the deep Q-learning network generates a perception combination of multiple vehicles to improve the perception synergy. Compared with the baselines, our proposed method improves the perception synergy. The experiment results show that our trained model has generalization ability and the end-to-end delay is under the constraint of safety critical applications.