Roadside Fisheye Vision for Cooperative Perception in V2I-Assisted Automated Driving

Precise road object perception and localization are crucial for autonomous vehicle navigation, yet onboard sensors occasionally encounter challenges with occlusions and blind spots, particularly at intersections. One potential solution is to use stationary sensors at intersections, which can enhance the perceptual capabilities of connected automated vehicles (CAVs) by leveraging vehicle-to-infrastructure (V2I) communication. In this context, this paper introduces an innovative perception and localization algorithm utilizing a stationary overhead fisheye camera installed at intersections. Addressing challenges inherent in overhead fisheye perspectives, a fine-tuning technique is employed to optimize detection performance for overhead traffic scenes. A novel camera calibration method is introduced to minimize localization inaccuracies derived from variations in road surface elevation. Road object dimensions are estimated for accurate localization and mapping in the birdeye view (BEV) map by fitting predefined 3D boxes in the real-world coordinate system. This is achieved by tracking and estimating object heading using the extended Kalman filter with the constant turn rate and velocity (CTRV) model. The proposed algorithm achieves remarkable localization accuracy, with a mean absolute error of 31 cm for pedestrians and 76 cm for cars, even at intersections with sloped roads. Experimental evaluations underscore the algorithms practical potential as a component for V2I-based cooperative perception and road safety warning systems.