Detection of Curb Boundary in Snow-covered Road Image via Feature Matching

Abstract

The automatic detection of snow-covered curb boundaries facilitates autonomous driving and assistance systems in enhancing safety and efficiency, particularly in scenarios wherein driving close to curbs is unavoidable, such as snow-removal vehicles operating near or regular vehicles approaching curbs. Most existing methods are heavily reliant on appearance-based and three-dimensional geometric features to identify curbs. These features may become ambiguous or vanish after heavy snowfall. However, even when curbs are completely obscured by snow, experienced drivers roughly estimate the curb boundaries for easy navigation. This ability relies on objects not covered by snow to assist with relative positioning. This study proposed an automated method to predict curb positions based on prior knowledge, mimicking the ability of experienced drivers to estimate curb boundaries even when obscured by snow. The proposed method leveraged Global Positioning System information to extract scenes that closely resembled current conditions from a pre-established database of snow-free scenes. By matching these scenes, the coordinates of the original curb boundaries were mapped onto snow-covered scenes to predict curb positions. Experiments on a circular route verified the effectiveness of our method. Furthermore, an evaluation metric was proposed to numerically assess the prediction results.