Hierarchical Temporal Convolution Network:Towards Privacy-Centric Activity Recognition

원문 초록

In response to the healthcare issues associated with the ageing population, various ambient assisted living technologies are being developed. To mitigate privacy concerns related to cloud-based data processing, recent methods have shifted towards using edge devices for local data processing. Despite their perceived benefits, the limited computational resources of these edge devices present a significant challenge for real-time performance, which is often an imperative requirement. However, recent computer vision-based methods for recognising activities of daily living among the elderly face increased computational complexity when capturing the multi-scale temporal context essential for accurate activity recognition. In this context, we propose HT-ConvNet (Hierarchical Temporal Convolution Network) to capture multi-scale temporal information without increasing computational complexity. HT-ConvNet employs exponentially increasing receptive fields across successive convolution layers to enable efficient hierarchical extraction of temporal features. Furthermore, HT-ConvNet provides an adaptive weighting mechanism to emphasise the most important features. Experimental results show that the multi-scale temporal feature extraction and the feature-weighted fusion mechanisms outperform existing methods in enhancing accuracy without increasing model complexity. The code is publicly available in: https://github.com/Gbouna/HT-ConvNet.