Integrated Visual Sensing and Efficient Deep Learning for UAV-Based Track Foreign Object Detection

Deploying deep learning models on UAV-based edge platforms for railway track foreign object detection is often limited by large model size and heavy computation. To address these issues, This paper presents an efficient foreign object detection model, termed RVGC-YOLO. The model introduces three key improvements: (a) designing a new C2f-RVG module to enhance multi-scale feature fusion; (b) replacing standard convolutions with GSConv to reduce parameters while maintaining feature extraction capability; and (c) embedding a coordinate attention mechanism to strengthen focus on critical regions. Furthermore, for efficient deployment on edge devices, this paper proposes an efficient model pruning and distillation framework, based on which a lightweight version of RVGC-YOLO, named RVGC-YOLO-Tiny, is introduced. Experiments on the UAV-RSOD dataset demonstrate that RVGC-YOLO achieves 89.60% mAP@0.5 and 63.72% mAP@0.5:0.95. Building upon this performance, the optimized RVGC-YOLO-Tiny significantly reduces parameters and computation by 61.86% and 55.06%, respectively, while maintaining a detection performance nearly consistent with the baseline. These results demonstrate that the proposed model optimization framework provides a practical solution for real-time railway safety monitoring.