关键词: 桥梁工程；裂缝分割；爬壁机器人；模块化设计；多尺度；实时轻量

Abstract: Aiming at the problems such as poor continuity of cracks on the surface of large-scale concrete bridge piers, large background interference, and large number of parameters of general deep learning crack detection model, a detection scheme combined with the improved lightweight DeepLabv3+ crack segmentation model and modular wall-climbing robot was proposed to achieve safe, fast and accurate detection of wall crack. The modular wall-climbing robot was used as the carrier to realize the crawling drive in complex environment through the self-group connection of each module, and the image acquisition equipment was equipped to collect the data of pier apparent disease. Meanwhile, based on DeepLabv3+ framework, a lightweight detection model of aggregation of multi-scale information was constructed by improving part of the network structure and adding various detection modules, and was deployed to the upper computer system. The final test results show that the average detection accuracy of the proposed model on the Crack-wall crack dataset reaches 86.96%, an improvement of 6.26% compared to the original model, an increase of 8.44% in intersection to union ratio, an increase of 8.76% in recall rate, and a model size of only 10.613M, with high detection accuracy and real-time detection effect. At the same time, the proposed detection scheme is feasible and successfully applied to the actual project.

Key words: bridge engineering; crack segmentation; wall-climbing robot; modular design; multi-scale; real-time lightweight