This paper introduces a railway track foreign body detection method based on the lightweight YOLO algorithm and the PP-LCNet network with integrated attention mechanism, effectively reducing the model size and enhancing the deployment performance and detection efficiency in resource-constrained environments. To address the issue of insufficient data in railway monitoring systems, a detection method utilizing TGA decomposition and semi-supervised learning is proposed to improve detection accuracy. To tackle the problem of sample imbalance, an unsupervised detection method based on the nonlinear characterization of deep local features is adopted, describing the high-dynamic background through auto-encoding embedding algorithms and constructing a sample core library to enhance detection efficiency. Ultimately, a detection method based on local feature anomaly scoring achieves efficient and reliable detection of track foreign bodies. The research holds significant practical value for railway traffic safety, effectively enhancing the efficiency and accuracy of foreign object detection in complex monitoring scenarios.