MARS: A Multiview Contrastive Approach to Human Activity Recognition From Accelerometer Sensor
Gulshan Sharma, Abhinav Dhall, Ramanathan Subramanian
Abstract
In this letter, we present <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">MARS</i> , a novel approach, which com-bines a multiview fusion technique with contrastive loss to accurately identify human activities using accelerometer sensor data. Accelerometer sensor enables precise monitoring of human activities in diverse contexts. Our approach leverages both temporal and spectral views of accelerometer data, integrating them through an attention mechanism to enhance the overall understanding of human activities. To further improve the discriminative power of the learned representations corresponding to different activity classes, we apply a contrastive loss-based siamese network. Emprical findings confirm that <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">MARS</i> outperforms state-of-the-art on the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">harAGE</i> dataset by a significant margin of 4.71 in unweighted average recall.