Intelligent detection technology for wall cracks at railway tunnel entrances based on UAV inspection

WU Kefeng; XU Guiyang; BAI Tangbo

doi:10.11860/j.issn.1673-0291.20240032

Abstract：Cracks in the wall of railway tunnel portals pose a serious threat to the operational safety of rail transport.To address the limitations of manual inspection,such as long inspection cycles and low accuracy,this study proposes a crack detection method for railway tunnel portal walls based on Un-manned Aerial Vehicle(UAV)inspection and the RFA-YOLOv8 model.First,high-resolution im-age data of railway tunnel portals are captured through UAV inspection.The collected images are pre-processed,and cracks in the tunnel portal wall are annotated to construct a training dataset for wall cracks.Second,the YOLOv8 object detection model is enhanced to better accommodate the characteris-tics of tunnel portal cracks.The Receptive Field Attention Convolution(RFAConv)is integrated into a newly designed C2f_RFA module,replacing the original C2f module in the backbone network to im-prove the model's focus on crack-prone areas.The BiFPN structure is introduced in place of the original feature fusion network to enhance the model's detection effect for targets of different scales.Addition-ally,the EIoU loss function is adopted to replace the CIoU,minimizing the differences in height and width between predicted box and ground-truth bounding boxes,thereby improving the model's detec-tion accuracy.Finally,the RFA-YOLOv8 model is validated and evaluated from three aspects:Com-parative experiments,ablation experiments,and visualization of detection results.Experimental results demonstrate that,compared to the original YOLOv8 model,the RFA-YOLOv8 model reduces the missed detection of small cracks,increasing recall by 3.8%and mean average precision by 2.5%.The proposed method effectively leverages UAV-captured tunnel portal images for accurate crack detection.