Improved YOLOX-based detection of condition of road manhole covers

Abstract

Manhole cover damage poses significant threats to road safety and infrastructure integrity, necessitating timely detection and repair. To address this, we introduce an enhanced YOLOX model integrated with ECA (High Efficiency Channel Attention) modules for real-time monitoring using car recorder footage. Our method categorizes manhole cover conditions into three distinct states: normal, broken, and down. By in-corporating ECA-Net before the decoupling head of the YOLOX model, we significantly boost its channel feature extraction abilities, critical for distinguishing subtle changes in cover conditions. Experimental results reveal a substantial increase in mean Average Precision (mAP) to 93.91%, with a notable AP of 92.2% achieved in the detection of the ‘down’ state, historically the most challenging category. Despite the en-hancements, our model maintains a high detection speed, processing at an average rate only five images per second slower than the original YOLOX model. Comparative analyses against leading detection models, in-cluding Faster R-CNN, SSD, and CenterNet, underscore the superiority of our approach in terms of both accuracy and speed, particularly in accurately recognizing the ‘down’ condition of manhole covers. This in-novative model provides a reliable tool for swiftly identifying damaged manhole covers and their precise lo-cations, enabling prompt maintenance actions. By improving the monitoring efficiency of urban infrastruc-ture, our solution contributes to enhanced road safety and the advancement of smart city technologies.