An Obstacle Intrusion Detection Method for Complex Environment Tracks Based on Manifold Regularization

Abstract

Tracks are essential infrastructure for high-speed railroads and are the key to the safe passage of trains. Once obstacles intrude into the track area, it will threaten the safety of train operations. Therefore, there is an urgent need to propose a fast and efficient obstacle intrusion detection method. However, it greatly challenges obstacle intrusion detection in the complex natural environment (rain, fog, light changes, etc.). To this end, this paper proposes an obstacle intrusion detection method for complex environment tracks based on manifold regularization. This method uses the single-stage object detection framework YOLO as the basic structure, integrates the multi-head attention mechanism to improve the object detection performance, and uses the dilated convolution structure to reduce the model parameters and improve the detection efficiency; in the feature extraction space, it introduces the manifold regularization constraint to realize the alignment constraint of various categories of features under different natural environment images and improves the generalization performance of the detection model. The results show that the proposed method can adapt to various complex natural environment detection conditions and effectively improve the accuracy and generalization of the obstacle intrusion detection method.