Multi-scale object detection for high-speed railway clearance intrusion

Clearance intrusion of foreign objects is a great threat to high-speed railway operation safety. An accurate and fast detection system of foreign object intrusion is important. In this paper, a multi-scale foreign object detection algorithm named feature fusion CenterNet with variable focus multi-scale augmentation (FFCN-VFMS) is proposed for the high-speed railway clearance scene. We render the ground truth with two-dimensional Gaussian distributions to generate the confidence score of each region in the image. In addition, a variable focus multi-scale augmentation (VFMS) method is proposed for multi-scale object detection, which takes detection results as prior knowledge to find the range of subsequent detection that contains most small objects. Moreover, feature fusion CenterNet (FFCN) adopts bidirectional iterative deep aggregation (BiIDA) to fuse the features in different convolutional layers and a spatial pyramid pooling (SPP) module to fuse feature maps extracted by different receptive fields. Our method was tested on public PASCAL VOC2007 datasets and our railway clearance intrusion (RCI) datasets. In comparison with related methods, FFCN-VFMS achieves better performance than comparison detectors with respect to accuracy and speed.