TeLeGaIT: Transfer Learning on Fog for Generalizable and Real-Time Transport Mode Detection

The knowledge of users’ transport modes, e.g., cars or buses, enables innovative transportation services, such as automatic ticketing and navigation. A promising approach for transport mode detection (TMD) is analyzing smartphone sensor data to infer transport modes using machine learning (ML) techniques. However, in existing work, the accuracy of MLbased approaches for a given city is highly dependent on whether the target ML model is trained with the data collected in that specific city by various users, smartphones, etc. To address this limitation, we need a comprehensive ML-based TMD system whose accuracy is not significantly affected by the geographical location, user variation, smartphone type, or smartphone position (e.g., in a bag or pocket)—generalizable TMD. In this paper, we propose an approach, called TELeGAIT, for generalizable TMD which is based on the idea of fine-tuning a base TMD ML model for the target city. Specifically, built on a base ML model trained with the aggregated data of some different cities, we fine-tune the base model for a new city by employing Transfer Learning (TL) and using the local data of that city. Moreover, some transportation services require low-latency TMD such as automatic ticketing. To this end, TeLeGait relies on a fog-based architecture, where the fine-tuning process is offloaded to a fog node in each city, and the TMD results are also stored on the fog node. TeLeGait utilizes a Multilayer Perceptron (MLP) model trained using two approaches: segment-based and raw sensor data. Our analysis shows that the raw data-based approach is more effective for neural networks, achieving $\mathbf{9 4. 4 2 \%}$ accuracy in detecting eight different transport modes. By employing TL, TELeGaiT generalizes the MLP model to improve the accuracy by $\mathbf{3. 6 \%}$ on average across different locations, users, and devices.