A Proposed Monocular Vision Method for Automated Paved Shoulder Detection in Videos

Abstract

The systematic recovery of design elements from the Brazilian federal highway network is vital for its efficient operation and maintenance. These information are necessary for studies that analyze the complete network or to prioritize interventions such as studies about on specific segments. This paper presents a computer vision algorithm for paved shoulder detection, using monocular cameras. Initially the lane and the possible shoulder are detected. Then the shoulder boundary is determined using length and homogeneity metrics. Finally, the shoulder on a frame is classified and, for error reduction, a sliding window is used to classify the shoulder in a point on the video. To evaluate the algorithm, videos of the federal highways BR-020 and BR-493 were used, achieving an accuracy of 89%.