Multi-stage activity inference for locomotion and transportation analytics of mobile users

In this paper, we, Ubi-NUTS Japan, introduce a multi-stage activity inference method that can recognize a user's mode of locomotion and transportation using mobile device sensors. We use the Sussex-Huawei Locomotion-Transportation (SHL) dataset to tackle the SHL recognition challenge, where the goal is to recognize 8 modes of locomotion and transportation (still, walk, run, bike, car, bus, train, and subway) activities from the inertial sensor data of a smartphone. We adopt a multi-stage approach where the 8 class classification problem is divided into multiple sub-problems considering the similarity of each activity. Multimodal sensor data collected from a mobile phone are inferred using a proposed pipeline that combines feature extraction and 4 different types of classifiers generated using the random forest algorithm. We evaluated our method using data from over 271 hours of daily activities of 1 participant and the 5-fold cross-validation. Evaluation results demonstrate that our method clearly recognizes the 8 types of activities with an average F1-score of 97%.