| With the continuous improvement of China’s transportation network,railway transportation has become the main artery of economic development.On the one hand,the popularity of high-speed rail makes railway transportation faster and faster.On the other hand,the number of newly built railways is increasing,which poses higher challenges to railway safety and railway maintenance.Fasteners,as an important part to fix the rails,are very important for the railway safety.The traditional method of replacing and repairing fasteners is to inspect the rail fasteners by workers walk along the road,which is inefficient and dangerous.With the development of machine vision,non-contact automatic detection methods are gradually applied in railway maintenance.At present,there are many studies on the detection of fastener missing domestic and overseas,but less on the location of fastener center.Generally,the overall positioning of fasteners is calculated roughly based on the relative geometrical relationship between fasteners and tracks and fasteners.However,this method is difficult to obtain the accurate position of fastener center.The central position of fasteners is provided to mechanical equipment,which can enable the automatic disassembly of fastener hexagonal nuts and further promote the automation of railway maintenance.The main work of this paper is to study the problem of fastener center location in image.The obtained pixel coordinates can be converted to the actual position in the world coordinate system by camera parameters and geometric relations.Aiming at the features of fastener image,this paper proposes a hierarchical positioning method from coarse to fine,which determinate the final fastener center positioning based on contour features.The main work of this paper is as follows:Firstly,the initial location of fasteners is done by using depth learning object detection,and the general position of fasteners is found from the whole image.On the one hand,the initial positioning of fasteners reduces the size of the image further processed and improves the efficiency of the algorithm;on the other hand,it reduces the interference of irrelevant factors and improves the positioning success rate and positioning accuracy.Deep learning object detection can be roughly divided into two kinds: R-CNN series algorithms based on candidate regions and YOLO series algorithms based on pairs.Considering the real-time performance,the faster YOLO algorithm is adopted in this paper.Then,aiming at the problem that the image of railway fasteners is easily disturbed by the surrounding environment,which leads to the unsatisfactory positioning accuracy of its geometric center,a non-classical receptive field interaction model inspired by biological vision is introduced to extract the contour of fasteners.The non-classical receptive field interaction model can preserve the real contour of the object and better the texture and noise in the image.According to the contour characteristics that need to be extracted,the butterfly non-classical receptive field model is improved,and the route non-classical receptive field model is proposed.The Canny operator based on the principle of non-classical receptive field suppression is introduced,and the iterative stimulus is introduced to modulate the gradient response,taking into account the efficiency and performance of the algorithm.Finally,Hough transform is used to extract the circle contour and hexagonal nut edge.Combining with the geometric constraints between the hexagonal nut edges,the real hexagonal vertexes are determined by voting,and the center position is located by these vertexes. |
