A Crosswalk Stripe Detection Model Based on Gradient Similarity Tags

Crosswalk detection is a significant component of intelligent driving systems. Recent deep learning models fail to detect crosswalk’s accurate shape. Traditional image processing techiniques are negatively affected by both irregular crosswalk stripe shapes and noises. In this paper, we present a novel crosswalk stripe detection model that reaches high accuracy and eliminates noises previous researches cannot remove. Two types of geometric features, the gradient similarity tags and classified vanishing points, are proposed to extract critical features and detect crosswalk stripes from a driver’s view. Our model can detect parallel crosswalk stripes, instead of a general crosswalk region, so that the detector can assist driving systems to make better decisions with pedestrian distribution, traffic signs, and other useful information. Our experimental results display high accuracy of detecting crosswalk stripes on the crosswalk image dataset, composed of crosswalk images captured at various road intersections. The images under different illumination conditions are detected correctly, which is not fixed by the previous Robust Zebra-crossing Detector and the former CROSSWATCH system. The method of computing the similarity coefficient, a parameter utilized to form similarity tags, contributes to crosswalk detection considerably. Multiple examples are presented to illustrate the process.