Exploiting Spatial and Temporal Features for Deep Learning Based Human Activity Recognition
Wenying Cao, Gaotao Shi, Tieguan Zhang · 2023
Human activity recognition using inertial sensors has gained significant popularity and widespread adoption in various fields, while deep learning has emerged as the dominant approach, playing a pivotal role in enhancing performance. Nevertheless, existing algorithms often neglect the heterogeneity of sensors and fail to effectively extract contextual features from long-term time series data, resulting in low accuracy in activity recognition, especially for complex activities. This paper proposes a novel method called FUsion Transformer hUman activity REcognition(FUTURE). FUTURE exploits the feature fusion mechanism and designs a multi-scaled DenseNet to extract the spatial features so that the relationship between heterogeneous sensors can be captured. Furthermore, a customized multi-head attention model with less computational complexity is employed in FUTURE to capture global dependencies within the sensor data. Experimental results on multiple datasets validate that FUTURE achieves an average recognition accuracy of 95% for complex activities and the evaluation against the state-of-the-art demonstrates the superior performance of the FUTURE model in accurately classifying complex activities.