N-Gram Based Transport Mode Detection Models for Energy Constrained Devices

Abstract

Transport mode detection is a subset of activity recognition. It is important for urban transportation planning and finding CO 2 footprint for individuals. However, daily usage of these applications are limited due to the power consumption of mobile devices. By the wide spread usage of wearable sensors, it will be crucial to develop systems that have smaller model size and less computational complexity. This paper presents a novel system that exploits n-grams which are widely used in natural language processing for extracting features of mobile sensor signals. This is the first study at the literature that provides smaller and faster activity recognition models by using n-grams. Our results demonstrate that, feature extraction with n-grams is less complex and creates 4 times smaller models even though its accuracy is comparable with classical methods. In addition, if n-gram based features and some of traditional features are combined, accuracy of the model achieves 89% which outperforms the research that use shallow machine learning approaches. Although this study focused on transport mode detection, we believe that n-gram based features could be used for several signal classification tasks.