Power-Aware Transport Mode Detection: A Comparative Analysis on Resource-Constrained Smartphones

Abstract

Understanding human mobility is crucial for applications like urban planning, traffic management, and personalized services. This paper presents a power-aware transport mode detection approach that leverages a limited set of smartphone sensors and machine learning algorithms to accurately classify various transportation modes while optimizing energy consumption. By focusing on a subset of essential sensors (accelerometers, gyroscopes, magnetometers, etc.) and implementing an optimized data preprocessing pipeline, our method reduces the energy drain associated with continuous data collection. Using the Sussex-Huawei Locomotion dataset, we evaluate seven classification models, namely, 1-Nearest Neighbor, Gaussian Naïve Bayes, Decision Tree, Random Forest, and two Convolutional Neural Networks and a ResNet model. We propose a preprocessing pipeline and a windowing strategy for temporal sensor data, and we demonstrate that Random Forest achieves high classification accuracy (95.05%). The study also discusses trade-offs between model performance, computational cost, and energy efficiency, highlighting the potential of well-designed lightweight models and sensor subsets for real-world deployment. Our findings suggest that traditional models, when properly optimized, can effectively support energy-efficient and privacy-conscious transport mode detection on smartphones.