Path voting based pavement crack detection from laser range images

Due to illumination variations, cast shadows, and pavement stains, etc., traditional optical imaging has limitations in capturing and representing pavement cracks. In this work, laser imaging techniques are used to model the pavement surface with point clouds, where crack points hold relatively lower range values than their non-crack neighbors. To extract cracks from laser range images, a two-level grouping approach is proposed. First, local grouping is performed by a novel segmentation-based path voting algorithm. The proposed path voting is equipped with an adapted normalized-cut algorithm which purposely bi-partitions an image patch along the potential crack path. Then in a global grouping, crack seeds are sampled and fed into a graph representation, in which spanning tree and tree pruning algorithms are employed to extract the final cracks in a global view. Experimental results demonstrate the effectiveness of the proposed approach.