铁路接触网支柱的图像序列自适应识别方法

Image Sequence Adaptive Recognition Method for Railway Catenary Pillar

摘要:

接触网支柱数字化管理是电气化铁路运维的关键环节,基于移动视频建立接触网支柱数字台账是高效、经济、便捷的技术手段。为实现对于移动视频图像序列中接触网支柱杆号的精准识别,提出了一种基于区域相关和改进SVTR网络的接触网支柱识别算法。针对视频图像中接触网支柱区域重叠、结构模式复杂的特点,采用了YOLO v4网络对单帧图像中支柱区域和号牌标识区域分别进行检测,并通过测算交叠区域来获得距观察点最近的杆位和对应的号牌区域。此外,针对接触网杆号牌尺度多样性和字符变长的问题,在杆号文字识别问题中采用了SVTR-tiny网络,并进一步引入迁移学习方法增强模型对于复杂杆号的识别精度和对于不同线路场景的泛化性能。通过在实际高铁线路采集的移动视频数据集上进行测试,结果表明算法在移动视频中视野最近杆位杆号区域的定位检出率可达98.01%,杆号文本的识别准确率达到96.13%,适用于我国高速铁路主要干线建设配套的接触网支柱结构。

Abstract:

Digital management of the contact network pillars is a critical component of operating and maintaining electrified railways. Creating a digital ledger of catenary pillars based on mobile video is an efficient, cost-effective, and convenient technological method. To achieve accurate recognition of the plate numbers on catenary pillars in mobile video image sequence, we propose a catenary pillar recognition algorithm based on region correlation and an improved SVTR network. Firstly, to address the challenges posed by overlapping contact network pillar areas and complex structural patterns in video images, we use the YOLO v4 network to separately detect catenary pillar areas and number plate areas in single-frame images. The nearest catenary pillar position and corresponding number plate area are then determined by calculating overlapping areas. Furthermore, we use an improved SVTR-tiny network for catenary pillar number text recognition to address the issues of diverse pillar number plate scales and variable character length, and transfer learning methods are introduced to enhance the model’s recognition accuracy for complex plate numbers and its generalization performance to different line scenes. Through testing on a mobile video dataset collected from actual high-speed railway lines, the test results show that the algorithm achieves a positioning recall rate of 98.01% for the nearest pillar number in the field of view and a pillar number text recognition accuracy of 96.13%, and the method is suitable for the catenary pillar structure supporting the construction of major high-speed railway trunk lines in China.