基于小样本学习的高铁接触网鸟窝检测方法研究

Research on High-Speed Rail Catenary Bird Nest Detection Method Based on Few-Shot Learning

摘要: 接触网作为电气化铁路供电系统的重要组成部分，承担着电力传输的重任，然而鸟窝的入侵可能造成接触网接地跳闸或绝缘击穿等故障，对行车安全造成危害。由于人工对鸟窝巡检的效率偏低，利用动车车载接触网运行状态检测装置(3C)采集的可见光全景图像，并基于深度学习方法来识别鸟窝已经成为当今技术发展的一种趋势。现有深度学习框架的训练方式大多是基于大量带标注的训练数据，而鸟窝在接触网中的稀缺性使得很难采集到充足的样本数据，且鸟窝的目标较小，容易受背景干扰。因此本文采用级联的方法对鸟窝进行识别，第一级检测网络基于YOLOv8算法，实现对鸟窝潜在区域(横梁、支柱和杆塔)的识别，排除大面积无关背景的干扰，第二级检测网络基于域适应小样本目标检测算法AcroFOD，以潜在区域作为输入，识别并定位鸟窝，通过域感知增强和定向优化策略，将源域的鸟窝特征迁移到目标域，达到96.9%的检出率和2.4%的误报率。此外，根据鸟窝区域的几何特征，对算法进行后处理优化，在保证相同检出率的基础上，误报率降低至0.1%，解决了工程化应用瓶颈。本文方法只需6张极少量的接触网鸟窝数据，就能实现鸟窝的有效检测，为后续接触网鸟窝的清除提供有力支撑。

Abstract: As a crucial component of the electrified railway power supply system, the catenary undertakes the important task of power transmission. However, the intrusion of bird’s nests may cause faults such as catenary ground tripping or insulation breakdown, posing risks to train operation safety. Due to the low efficiency of manual bird’s nest inspection, using visible light panoramic images collected by the EMU-mounted catenary operation status detection device (3C) and identifying bird’s nests based on deep learning methods has become a trend in current technological development. Most training methods of existing deep learning frameworks rely on a large amount of annotated training data. Nevertheless, the scarcity of bird’s nests in the catenary makes it difficult to collect sufficient sample data; additionally, bird’s nests are small targets and easily interfered by the background. Therefore, this paper adopts a cascaded approach to identify bird’s nests. The first-stage detection network is based on the YOLOv8 algorithm, which realizes the identification of potential bird’s nest areas (beams, pillars, and poles) and eliminates interference from large areas of irrelevant backgrounds. The second-stage detection network is based on AcroFOD, a domain-adaptive few-shot object detection algorithm. It takes potential areas as input to identify and locate bird’s nests. Through domain-aware enhancement and directional optimization strategies, the bird’s nest features from the source domain are transferred to the target domain, achieving a detection rate of 96.9% and a false alarm rate of 2.4%. Furthermore, according to the geometric features of bird’s nest areas, post-processing optimization is performed on the algorithm. On the basis of ensuring the same detection rate, the false alarm rate is reduced to 0.1%, solving the bottleneck in engineering applications. The method proposed in this paper only requires a very small number of catenary bird’s nest data (6 samples) to achieve effective detection of bird’s nests, providing strong support for the subsequent removal of catenary bird’s nests.