Contrastive-representation IMU-based fitness activity recognition enhanced by bio-impedance sensing

Abstract

While IMU-based Human Activity Recognition (HAR) has achieved significant success in wearable and pervasive computing areas over the past decade, the potential for further improvement of IMU-based HAR performance through the contrastive representation method enhanced by other sensing modalities remains underexplored. In this work, we propose a contrastive representation learning framework to demonstrate that bio-impedance can enhance IMU-based fitness activity recognition beyond the common sensor fusion method, which requires all sensing modalities to be available during both training and inference phases. Instead, in our proposed framework, only the target sensing modality (IMU) is required at inference time. To evaluate our method, we collected both IMU and bio-impedance sensing data through an experiment involving ten subjects performing six types of upper-body and four kinds of lower-body exercises over five days. The bio-impedance-alone classification model achieved an average Macro F1 score of 75.49% and 71.57% for upper-body and lower-body fitness activities, respectively, which was lower than that of the IMU-alone model (83.10% and 78.61%). However, with our proposed method, significant performance improvement (2.66% for upper-body activities and 3.2% for lower-body activities) was achieved by the IMU-only classification model. This improvement leverages the contrastive representation learning framework, where the information from bio-impedance sensing guides the training procedure of the IMU-only model. The results highlight the potential of contrastive representation learning as a valuable tool for advancing fitness activity recognition, with bio-impedance playing a pivotal role in augmenting the capabilities of IMU-based systems.