Robust Zebra-Crossing Detection for Autonomous Land Vehicles and Driving Assistance Systems

Abstract

Road scene understanding is critical for driving assistance systems and autonomous land vehicles. The main function of road scene understanding is robustly detecting useful visual objects existing in a road scene. A zebra crossing is a typical pedestrian crossing used in many countries around the world. When detecting a zebra crossing, an autonomous lane vehicle is normally required to automatically slow down its speed and to trigger a path-planning strategy for passing the zebra crossing. Also, most of driving assistance systems can send an early-warning signal to remind drivers to be more careful. This paper proposes a robust zebra-crossing detection algorithm for autonomous land vehicles and driving assistance systems. Firstly, an inverse perspective map is generated by utilizing camera calibration parameters to obtain a bird-eye view road image. Secondly, a course-to-fine detection process is applied to obtain a candidate zebra-crossing region and finally a true zebra-crossing region is recognized by combining appearance and shape features. Experiments on several kinds of real road videos which also include several challenge scenes demonstrate the effectiveness and efficiency of the proposed method.