An Pavement Crack Detection Method Based on Edge Computing Platform

This paper proposes a computer vision system for pavement crack detection and an image segmentation algorithm applicable to the system. Conventional vehicle-mounted pavement crack detection systems are based on bulky mainframe or computer architectures, whereas we develop a more lightweight one. The system consists of a single camera, a small light source and an edge computing platform that can be easily installed on any carrier. Our lightweight segmentation algorithm for this system has an asymmetric encoder-decoder structure. The modified inverse residual module for fast downsampling is designed for the encoder, the lightweight contextual information extraction module is designed for the decoder, and the segmentation prediction results are obtained by fusing shallow feature maps. The model is trained on a self-built dataset and accelerated for the system's edge computing platform. Experiments show that our system can complete the pavement crack detection of the captured images with a single frame runtime of 0.0209s and 47.8FPS. And it achieves 98.10% pixel segmentation accuracy and 76.54% mIoU, showing a competitive performance.