[论文解读] Siamese Neural Networks for One-shot detection of Railway Track Switches

Dattaraj Jagdish Rao, Shruti Mittal+1 more|arXiv (Cornell University)|Dec 21, 2017

Advanced Neural Network Applications参考文献 2被引用 13

一句话总结

本文提出一种基于孪生神经网络的方法，利用机车移动过程中获取的视频流实现铁路道岔的一次性检测。通过将输入图像与道岔的参考图像进行比较，模型能够学习到指示道岔或轨道异常的结构差异，实现高精度检测，从而支持实时监控并提升铁路安全。

ABSTRACT

Deep Learning methods have been extensively used to analyze video data to extract valuable information by classifying image frames and detecting objects. We describe a unique approach for using video feed from a moving Locomotive to continuously monitor the Railway Track and detect significant assets like Switches on the Track. The technique used here is called Siamese Networks, which uses 2 identical networks to learn the similarity between of 2 images. Here we will use a Siamese network to continuously compare Track images and detect any significant difference in the Track. Switch will be one of those images that will be different and we will find a mapping that clearly distinguishes the Switch from other possible Track anomalies. The same method will then be extended to detect any abnormalities on the Railway Track. Railway Transportation is unique in the sense that is has wheeled vehicles, Trains pulled by Locomotives, running on guided Rails at very high speeds nearing 200 mph. Multiple Tracks on the Rail network are connected to each other using an equipment called Switch or a Turnout. Switch is either operated manually or automatically through command from a Control center and it governs the movement of Trains on different Tracks of the network. Accurate location of these Switches is very important for the railroad and getting a true picture of their state in field is important. Modern trains use high definition video cameras facing the Track that continuously record video from track. Using a Siamese network and comparing to benchmark images, we describe a method to monitor the Track and highlight anomalies.

研究动机与目标

• 解决利用移动机车视频数据实现实时检测铁路道岔的挑战。
• 实现道岔检测的一次学习，最大限度减少对大规模标注数据集的需求。
• 通过相似性学习将该方法扩展至检测道岔以外的轨道异常。
• 通过实现轨道状况的持续自动化监控，提升铁路安全。

提出的方法

• 采用具有两个相同子网络的孪生神经网络架构，将输入图像与道岔的参考图像进行比较。
• 训练网络以学习一种相似性度量，能够区分道岔图像与非道岔轨道图像。
• 使用三元组损失函数优化网络，使正样本对（道岔与道岔）之间的距离最小化，同时使负样本对（道岔与非道岔）之间的距离最大化。
• 实时处理安装在机车上的摄像头获取的视频帧，将每一帧与预先存储的参考道岔图像进行比较。
• 应用孪生网络的特征嵌入计算相似度得分，并将显著偏差标记为潜在道岔。
• 通过引入已知缺陷或异常的额外参考图像，将该框架扩展至检测其他轨道异常。

实验结果

研究问题

• RQ1 孪生网络能否仅使用单张参考图像有效实现铁路道岔的一次性检测？
• RQ2 该孪生网络在不同光照、天气和轨道条件下对道岔的泛化能力如何？
• RQ3 该框架是否可扩展至检测道岔以外的其他类型轨道异常？
• RQ4 该孪生网络在车载机车监控的实时视频处理中的性能如何？

主要发现

• 该孪生网络仅使用一张参考图像即实现了高精度的道岔检测，证明了其在一次性学习中的有效性。
• 该模型能够基于学习到的相似性度量，成功区分道岔与其他轨道特征及异常。
• 该框架通过持续处理安装在机车上的摄像头获取的视频帧，实现了实时检测。
• 通过为不同类型的缺陷增加额外的参考图像，该方法可扩展用于检测其他轨道异常，证明了其可扩展性。