A Real Time Bird Nest Detector for Railway Catenary Safety Inspection
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
With the rapid development of China’s high-speed railway network, contact network online inspection based on on- board video has become an essential method for ensuring the safety of the power supply system. However, existing detection models generally face challenges such as insufficient accuracy, limited inference speed, and large model size. To address these issues, this paper proposes a lightweight detection framework, RailVNet, based on YOLOv8. First, a contact network tower nest dataset, RailV-VID, was collected and constructed, containing various train speeds, lighting conditions, and background scenarios. Secondly, in the model design, DS-ShuffleNetV2 (Dynamic-Shuffle- ShuffleNetV2) was used to replace the backbone, reducing the model size to 1.9M and significantly lowering com- putational costs. In the Neck section, AW-BiFPN (Adaptive-Weighted BiFPN) is introduced, leveraging bidirectional feature flow and the ”dual-input weighted Concat + normalization” mechanism to efficiently complement multi-scale semantics and shallow textures. Finally, an RCA module (Rail-Context Attention, based on CBAM and deformable convolution fusion) is embedded at the Head end, enhancing the perception of small targets and irregular nest struc- tures in complex backgrounds. Experiments show that RailVNet achieves mAP50 = 92.9% and mAP50-95 = 52.6% on the RailV-VID dataset, improving by 12.3% and 24.6%, respectively, compared to the baseline model. At the same time, the number of parameters is reduced by 83.0%, and the inference speed reaches 145.4 FPS. This framework balances both high precision and real-time performance, providing an engineered, lightweight solution for on-board contact network inspection, with broad application prospects.