Train detection from audio using one class classification
Electronics Engineering, MSc. Thesis, 2015
Assoc. Prof. Dr. Hakan Erdogan (Thesis Advisor),
Assoc. Prof. Dr. Berrin Yanikoglu ,
Assoc. Prof. Dr. Erchan Aptoula.
Date & Time: 24th December 2015 – 4:00 PM
Place: FENS L065
Keywords : Audio source recognition, MFCC, One class classifier, SVDD
In this thesis, we focus on detecting a train from the sound generated by it. An audio sensor is placed close to a railway track to record ambient sounds which may or may not originate from a train. In this problem, we define the target event as the recording of a train sound and outlier events are all other audio events that are recorded by the audio sensor.
In machine learning and pattern recognition, classifiers are trained from labeled data to categorize a new observation. Classifiers are usually trained from data which contain all possible classes, however it is possible that during training the classifier, for some classes the data is either not available or it is so diverse in nature that it cannot be used reliably. In case of binary classification, if one of the classes do not have reliable training data, we can use a "one class classification" strategy which only uses a single class data for training.
For train detection from audio, we compared a one class classifier called support vector data description (SVDD) with binary classifiers and showed that SVDD performs well in cases where data from the outlier class is scarce. We also tested the SVDD trained model in real time and the results indicate that the goal of reducing the false positive rate is satisfactorily achieved. The tests are performed using audio data recorded in Bathmen, a town in eastern Netherlands, by the company Sensornet for a project about train detection and sound level monitoring.