Pavement Crack Detection on BEV Based on Attention-Unet

Abstract

Identifying and detecting pavement cracks quickly and accurately for traffic safety is one of the important problems in the field of automatic driving. This study presents a framework of crack detection on BEV (Bird's Eye View). Firstly, based on the binocular parallax information, the captured road image is transformed from perspective to BEV as the input of the network. The Unet with attention mechanism is used to selectively fuse the deep and shallow features to identify the cracks on the pavement. In addition, further processing is performed according to the results of crack detection. The results help judge the quality of the pavement and provide a basis for the measurement of crack width in the direction of normal vector, laying a foundation for subsequent application. The test shows the method has high detection accuracy and is suitable for complex pavement conditions.