Road Meteorological State Recognition in Extreme Weather Based on an Improved Mask-RCNN

Guangyuan Pan, Zhiyuan Bai, Liping Fu, Lin Zhao, Qingguo Xiao

Abstract: Road surface condition (RSC) is an important indicator for road maintenance departments to survey, inspect, clean, and repair roads. The number of traffic accidents can increase dramatically in winter or during seasonal changes when extreme weather often occurs. To achieve real-time and automatic RSC monitoring, this paper first proposes an improved Mask-RCNN model based on Swin Transformer and path aggregation feature pyramid network (PAFPN) as the backbone network. A dynamic head is then adopted as the detection network. Meanwhile, transfer learning is used to reduce training time, and data enhancement and multiscale training are applied to achieve better performance. In the first experiment, a real-world RSC dataset collected from Ministry of Transportation Ontario, Canada is used, and the testing result show that the reidentification accuracy of the proposed model is superior to that of other popular methods, such as traditional Mask-RCNN, RetinaNet, Swin Double head RCNN, and Cascade Swin-RCNN, in terms of recognition accuracy and training speed. Moreover, this paper also designs a second experiment and proved that the proposed model can accurately detect road surface areas when light condition is poor, such as night time in extreme weather.