Efficient Human Activity Recognition Based on Grouped Representations of Multimodal Wearable Data

Human Activity Recognition (HAR) is a vast and complex research domain that has multiple applications, such as healthcare, surveillance or human-computer interaction. Several sensing technologies exist to record data later used to recognize people’s activity. This paper aims to linger over the specific case of HAR based on multimodal wearable sensing devices. Corresponding HAR datasets provide multiple sensors information collected from different body parts. Previous approaches consider each information separately or altogether. Vision HAR methods consider each body segment and their position in space in order to perform activity recognition. This paper proposes a similar approach for Multimodal Wearable HAR (MW-HAR). Datasets are first re-sampled at a higher sampling rate (i.e., lower frequency) in order to both decrease the overall processing time and facilitate interpretability. Then, we propose to group sensing features from all the sensors corresponding to the same body part. For each group, the proposal determines a different representation realm of the group information. This abstracted representation depicts the different states of the corresponding body part. Finally, activity recognition is performed based on these trained abstractions of each considered body part. We tested our proposal on three benchmark datasets. Our evaluations first confirmed that a re-sampled dataset offers similar or even better performance for activity recognition than usual processing. But the primary advantage is to decrease significantly the training time. Finally, results show that a grouped abstraction of the sensors features is improving the activity recognition in most cases, without increasing training time.