Crack segmentation-guided measurement with lightweight distillation network on edge device

Abstract

Pavement crack measurement (PCM) is essential for automated, precise road condition assessment. However, balancing speed and accuracy on edge artificial intelligence (AI) mobile devices remains challenging. This paper proposes a real-time PCM framework for edge deployment, incorporating a lightweight distillation network and a surface feature measurement algorithm. Specifically, the proposed instance-aware hybrid distillation module combines feature-based and relation-based knowledge distillation, leveraging crack instance-related information for efficient knowledge transfer from teacher to student networks, which results in a more accurate and lightweight segmentation model. Additionally, a real-time crack surface feature measurement algorithm, based on distance mapping relationships and crack edge coordinate extraction, addresses issues with crack edge branching and loss, enhancing measurement efficiency. Real-time measurement was performed on actual roads utilizing mobile robot equipped with an edge computing unit. The crack segmentation precision reached 84.37%, with a frame per second of 77.72. Compared to the ground truth, the relative error for average crack width ranged from 6.42% to 40.65%, while the relative error for crack length varied between 1.48% and 3.76%. These findings highlight the feasibility of real-time crack assessment and save road maintenance costs.