A Convolved Self-Attention Model for IMU-based Gait Detection and Human Activity Recognition

This paper presents a convolved self-attention neural network model for gait detection and human activity recognition (HAR) tasks using wearable inertial measurement unit (IMU) sensors. By embedding a convolved window inside the self-attention module, prior time step knowledge is utilized by self-attention layer to improve accuracy. Moreover, a streamlined fully connected (FC) layer without hidden layers is proposed for the feature mixer. This arrangement enables significant reduction of overall network parameters, since hidden layers occupy the majority of the parameters in a transformer encoder. Compared to the other state-of-art neural networks, the proposed method achieved better accuracy of 95.83% and 96.01% with the smallest network size on HAR datasets UCI-HAR and MHEALTH respectively,