The integrity of railway fences is critical to the safe operation of railways, with fence damage and unauthorized intrusions posing significant safety threats. In recent years, drone-based railway fence monitoring technology has emerged as a research focus. However, traditional methods struggle to effectively detect fence damage in complex railway scenarios, particularly due to insufficient accuracy in small object detection and personnel intrusion identification under complex background interference. To address this, this study proposes a railway fence damage and personnel intrusion detection method that integrates SAM2 and an improved YOLOv8. SAM2 is utilized for high-precision semantic segmentation of fence areas, while the improved YOLOv8 is employed for detecting intruders within the fenced regions. The enhanced YOLOv8 incorporates a RepVGG reparameterization convolution module, which optimizes 3×3 and txt convolutions in parallel during training and merges them during inference, significantly improving feature extraction capabilities. Additionally, a CBAM convolutional attention module is embedded in the backbone network to enhance the specificity of feature representation. A new 160×160 small object detection branch (P2) is introduced, markedly improving small object detection performance. Experimental results demonstrate that the proposed network achieves a detection accuracy of 82.7% for small object detection tasks, significantly surpassing traditional methods in fence damage detection and personnel intrusion identification from drone perspectives, thus providing reliable technical support for railway safety management.
Research on Intrusion Detection Based on SAM2 and Improved YOLOv8
13.06.2025
2449557 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch