Robust and Resilient Cooperative Perception under V2X Communication Limitations

Abstract: Environmental perception is fundamental to safe and efficient autonomous driving. With Cooperative Perception (CP) enabled by V2X networks, connected vehicles can exchange perceptual information to see through blind zones and deal with long-tail scenarios. In this talk, we propose a robust, reliable, and resilient CP framework for connected autonomous driving under V2X Communication Limitations. First, for robustness to localization error and communication delay, a calibration-free two-stage CP paradigm is proposed using deep metric learning. This fusion method only requires image data and is adaptive to the transmission rate. Then, to guarantee high reliability, hard AoI constraints are considered in sensor scheduling of CP to guarantee the timeliness of perceptual information. The required channel resources are minimized in asynchronous status update settings. Next, to resiliently adapt to the dynamic traffic environment, we propose a learning-while-scheduling approach to trade off exploration and exploitation. An online sensor scheduling algorithm is designed based on restless MAB (Multi-Armed Bandit) theory to maximize the average CP gain with low scheduling overhead. Finally, a large-scale multi-view multi-modality dataset, called Dolphins, is presented to assist further researches and verification of CP <a href="http://systems.

Room:” target=”_blank” title=”systems.

Room:”>systems.

Room: 660, Bldg: ECS , University of Victoria, Victoria, British Columbia, Canada, V8P5C2