Robust and Real-time Road Crack Detection through Collaborative Dual-Branch Learning on Robotic Sensing Platform

In the emerging field of urban digital twins, the development of intelligent road inspection robots and automated crack detection systems plays a crucial role in infrastructure maintenance. Although deep learning-based pavement crack detection algorithms have shown promising results, they are predominantly designed for images captured at close range. As the distance between the crack and the imaging device increases, particularly in the presence of background interferences like zebra crossings, these algorithms often experience a significant decline in performance. To address this issue, we propose DualCrackNet, an innovative network that integrates both localization and segmentation capabilities for robust road crack detection. Experimental results demonstrate that, through collaborative dual-branch training and the designed boundary extraction module, DualCrackNet outperforms existing road crack detection algorithms on both our proposed RCD900 dataset, which contains background interferences, and the publicly available DeepCrack dataset. Also, DualCrackNet surpasses these algorithms in processing efficiency with 239.55 frame per second (FPS) on a single NVIDIA RTX3090. In addition, we present CrackRobo, a robotic platform designed to provide a feasible technical pathway for the next generation of intelligent road inspection robots. The source code package and proposed RCD900 dataset are available at https://github.com/blackspiderrr/DualCrackNet.