Human Activity Analysis in Smart Environments

Abstract

Human Activity Recognition (HAR) in smart environments is central to Ambient Assisted Living (AAL), which focuses on supporting independent living, safety, and well-being for older adults and individuals requiring assistance. Vision-based HAR systems using cameras capture rich contextual data but face limited user acceptance due to privacy concerns. This thesis presents contributions to the design of vision-based HAR systems aimed at addressing end users’ concerns to engender trust and enhance acceptance. The first contribution addresses the lack of empirical evidence on user perspectives on vision-based HAR systems. An online survey explored the perspectives of prospective users, identifying several design requirements such as communication and transparency, system flexibility and output modes, reliability and interpretability, privacy and data security, user control, and clarity and simplicity. The second contribution presents skeleton-based HAR as a privacy-preserving alternative to RGB data. Two models are proposed. The Hierarchical Temporal Convolution Network (HT-ConvNet) introduces feature-weighted fusion and a method for capturing long-range temporal dependencies without added computational cost, improving both accuracy and efficiency. The Robust and Efficient Temporal Convolution Network (RE-TCN) combines adaptive temporal weighting, depthwise separable convolutions, and data augmentation to enhance robustness and efficiency in unstructured real-world settings. Both models outperform existing approaches on multiple datasets. The third contribution explores millimetre-wave (mmWave) radar as an alternative sensing modality. The proposed Stochastic Edge Network (SEdgeNet) enables direct activity recognition from sparse point clouds using adaptive neighbour sampling and efficient edge feature learning. Complementing this, the proposed mmPoint-PoseNet estimates 3D human poses from mmWave radar point clouds, providing 3D pose data to support downstream HAR. Both methods achieve state-of-the-art results on activity recognition and pose estimation, addressing the challenges of noise, sparsity, and irregularity in mmWave radar data. These contributions advance the design of HAR systems in smart environments by aligning technical development with user expectations. The work demonstrates how AAL systems can be designed to foster trust and enhance adoption.