Abnormal Activity Detection Based on Place and Occasion in Virtual Home Environments

This paper focuses on enhancing the safety of older adults in their home environment by analyzing and distinguishing between abnormal and normal states in their daily living activities. We initially propose simulating virtual abnormal and everyday activities to address the challenge of limited real-world datasets containing abnormal activities. The technical approach consists of three main components to achieve our purpose: object detection, feature extraction, and a context-aware decision-making process. Specifically, object detection is performed using transfer learning from the YOLO pre-trained model to identify objects within the environment. Then, we propose a virtual grounding point feature extracted from skeleton images, enabling the prediction of transition from a normal to an abnormal human posture. Furthermore, the likelihood of posture deformities is calculated using skeleton joint points. Finally, the decision-making process takes into account the place and occasion within the home environment by using the Hidden Markov Model to provide the abnormal and normal state discrimination for context-aware safety assessments. Our proposed approach utilizes the VirtualHome2KG dataset, which has demonstrated its effectiveness in identifying abnormal and normal situations to enhance the safety of older adults.