Semi-supervised Segmentation of Accelerometer Time Series for Transport Mode Classification

Collecting ground truth data with smart phone applications is as difficult as important for training classification models predicting transport modes of people. Errors of respondent input with respect to trip length and transport mode segmenting introduce a systematic bias in the classification model. We propose a semi-supervised framework adjusting user given input to process user-collected accelerometer time series data. Our contributions are (1) an evaluation of the impact of segmentation bias, (2) a novel algorithm to find more homogeneous segments and (3) a robust incrementally trained classifier model based on clustering employing Dynamic Time Warping as similarity measure. We apply the proposed method on synthetic and real-world accelerometer trip data of 800 labeled trips consisting of 2000 user-given segments and 400 hours travel time and test it against a baseline classifier relying completely on user-feedback. The results prove that our method learns clusters revised from noise and increases the classifier's accuracy for real world and synthetic data by up to 17%.