Intelligent Crack Detection in Infrastructure Using Computer Vision at the Edge

Abstract

To fulfil the demands of the industry in autonomous intelligent inspection, innovative frameworks that allow Convolutional Neural Networks to run at the edge in real-time are required. This paper proposes an end-to-end approach and system to enable crack detection onboard a customised embedded system. In order to make possible the deployment and execution on edge, this work develops a dataset by combining new and existing images, it introduces a quantization approach that includes inference optimization, memory reuse, and freezing layers. Real-time, onsite results from aerial and hand-held setup images of industrial environments show that the system is capable of identifying and localiszing cracks within the field of view of the camera with a mean average precision (mAP) of 98.44% and at ~2.5 frames per second with real-time inference. Therefore, it is evidenced that, despite using a full model, the introduced model customization improved the mAP by ~8% with respect to lighter state-of-the-art models, and the quantization technique led to a model inference two times faster. The proposed intelligent and autonomous approach advances common offline inspection techniques to enable on-site, artificial intelligence-based inspection systems, which also aid in reducing human errors and enhance safety conditions by automatically performing defect-recognition in tight and difficult-to-reach spots.