Key points are not available for this paper at this time.
Human activity recognition (HAR) utilizing motion data from inertial sensors embedded in wearable devices plays a crucial role in healthcare monitoring, fitness assessment, and ambient assisted living systems. Nevertheless, the precise classification of complex activities remains a significant challenge due to signal noise, individual variability, and the temporal intricacy of human motion. This study introduces Residual-BiGRUNet, a novel deep learning framework integrating bidirectional gated recurrent units (BiGRU) with residual connections. The architecture addresses vanishing gradient issues and captures bidirectional temporal patterns within sequential sensor data. Evaluation was conducted on the WISDM benchmark dataset, comprising accelerometer data from six everyday daily activities performed by 36 participants. Using Leave-One-Subject-Out cross-validation, the proposed model achieved an accuracy of 99.55%, precision of 99.21%, recall of 99.79%, and F1-score of 99.49%, outperforming conventional BiGRU models by 1.53 percentage points. Including residual pathways enhances feature representation and enables more effective differentiation between similar activities such as walking, jogging, and stair climbing. These results indicate that incorporating residual learning into bidirectional recurrent networks holds strong potential for accurate and efficient HAR in real-world, resource-constrained settings.
Fan et al. (Fri,) studied this question.