Synthetic Ground Truth Data Generation for Automatic Trajectory-based ADL Detection

In-house automatic activity detection is highly important toward the automatic evaluation of the resident's cognitive state. However, current activity detection systems suffer from the demand for on-site acquisition of large amounts of ground truth data for training purposes, which poses a major obstacle to their real-world applicability. In this paper, focusing on resident location trajectory-based activity recognition through limited amount of low-cost cameras, we introduce a novel scheme for automatic ground truth data generation, via simulation of resident trajectories based on formal descriptions of activities. Additionally, we present an activity detection scheme capable of learning activity patterns from such synthetic ground truth data. Experimental results show that our methodology achieves activity detection performance that is comparable to state-of-art methods, while suppressing the need for any actual ground truth recordings, thus boosting the real-world applicability of practical activity detection systems.