UTILIZING SCANPATH TREND ANALYSIS FOR EFFICIENT REAL-TIME PRIVACY-PRESERVING HUMAN ACTIVITY RECOGNITION WITH WEARABLE SENSORS

Many of the state-of-the-art human activity recognition (HAR) systems rely on computationally demanding models for training and recognition, which require healthy communication to remote locations such as cloud services. They also require the sharing of personal data across multiple devices but data sharing is discouraged, since privacy is becoming an important concern. In this thesis, the Scanpath Trend Analysis (STA) algorithm is used for the first time with inertial sensors to achieve privacy-preserving HAR system. With PAMAP2 dataset, for intra-subject, the proposed approach performs better than the existing studies with 91.8% accuracy and for inter-subject recognition, it has comparable accuracy of 92.0%. To evaluate the proposed approachs efficiency and edge computability, a testbed is introduced and the system is implemented on a Raspberry Pi 3B and compared with decision tree-based approaches which are known for CPU and memory efficiency. The proposed approach performs better in terms of computation time for training, smaller peak memory usage and model sizes. These findings suggest that the proposed approach can achieve accurate and efficient HAR on edge devices while maintaining privacy.