Robust Learning for Autonomous Driving Perception Tasks in Cyber-Physical-Social Systems

Abstract

Modern data-driven perception systems accomplish excellent performance in a simple, constrained environment but have unreliable detection in real, complex, and challenging weather situations. Therefore, robust perception performance in new and unanticipated domains is a crucial factor for autonomous driving. To this end, this paper proposes a novel robust learning scheme for arbitrary domain perturbations. In this paper, the model robustness is characterized by the anchor data and corresponding various domain shift directions. The robust learning problem is then formulated as a min-max optimization problem conjugated to the constrained exploration space defined by the perturbation model and the semantic parameter. The proposed robustify procedure solves the optimization problem considering the worst-case scenario, which improves the model resilience in multiple domain shift directions, especially for those variations in Cyber-Physical-Social Systems (CPSS).