Automatic Detection of Train Whistle

• Main Authors: LIN, YAN-TING, 林彥廷
• Other Authors: TSAI, WEI-HO
• Format: Others
• Language: zh-TW
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/pncu8t

碩士 === 國立臺北科技大學 === 電子工程系 === 107 === This work studies the methods of automatically detecting train whistles, which could serve as a warning assistance for railway workers or avoid accidents arising from the unexpected abnormality of the railroad crossing signal. Our basic strategy is to distinguish train whistles from other sounds that may occur in the area near the railway, such as wind sound, rain sound, and wave sound. However, as the recorded train whistle signals may be intermixed with other sounds, how to distinguish train whistles from other sounds is a challenging problem. This thesis investigates the three methods, namely, support vector machine (SVM), convolutional neural network (CNN), and long term short memory neural network (LTSM). Our experiments conducted using both the data from websites and recorded by ourselves show that SVM performs the best, CNN performs the worst, and LSTM are between these two extremes. The classification accuracies are 95.8%, 87.0%, and 89.2%, respectively.