Abstract: Unstructured road potholes in open pit mining areas have always been one of the important problems plaguing
the safety of unmanned mining cards. In order to solve the safety problems caused by potholes,such as the bumpy and even
rollouts of unmanned mining cards,a YOLOv7-RFEM model for road pothole detection in open pit mining areas was proposed
in this paper. Firstly,for potholes with different sizes,irregular shapes and not obvious features,a Scale-Aware RFE module
based on cavity convolution and shared weights is introduced into Neck to further expand the sensitivity field of feature maps
and improve the multi-scale detection performance and accuracy of the model while reducing the number of parameters. Secondly,
in view of the actual situation of the integration of road background and pothole features,the model adds an Efficient Multi-
Scale Attention mechanism (EMA) for cross-space learning to ELAN,which enhances the characteristics of small targets of
potholes and weakens the background interference of potholes. Finally,to solve the problem of inaccurate location of the pothole
boundary frame,CIoU Loss was replaced by NWD Loss,so that the model paid more attention to the overlap between the predicted
frame and the real frame,and adapted to the changes of the shape and size of the pothole. The experiment of a large
opencast mine in Hami,Xinjiang proves that the model has good detection effect in many cases,and can meet the demand of
accurate detection of unmanned mining cards in opencast mine.

Key words: road pothole detection,unmanned driving,feature extraction,safety warning