Traffic Obstacles Comprehensive Identification Using Shape Descriptors in Video

Abstract

Traffic obstacle identification has been widely studied as part of the broader obstacle detection area for Autonomous Vehicles (AV). Existing in-vehicle perceptual systems are concentrated on obstacle detection for pedestrian or vehicle, and limited work has been conducted on obstacle comprehensive identification. In general, the feature selection is particularly critical in the process of obstacle identification. As a set of features to describe a given shape or contour, shape descriptor have attracted much attention in recent years and play an important roles in target recognition. This paper proposed a shape descriptor based obstacle comprehensive identification method where the traffic obstacles be classified into four classes: vehicle, lateral moving pedestrian, longitudinal moving pedestrian, and unknown (such as trees, road lamp etc.). Here first a variety of shape descriptors be extracted from the contour curve, such as Hu moment, Zernike moments, Krawtchouk moments, Shape Context, and Axis of Least Inertia. Then, these shape descriptors be used to identification obstacles respectively. Finally, the identification results are contrastive analyzed. Though a single shape descriptor does not achieve ideal identification results for traffic obstacle, but this will provide a new idea for obstacle comprehensive identification using shape descriptor in video for AV.