Lifelogging system based on averaged Hidden Markov Models: dangerous activities recognition for caregiver support

In this paper, a prototype lifelogging system for monitoring people with cognitive disabilities and elderly people as well as a method for the automatic detection of dangerous activities are presented. The system allows for the remote monitoring of observed people via an Internet website and respects the privacy of the people by displaying their silhouettes instead of their actual images. The application allows for the viewing of both real-time and historical data. The lifelogging data (skeleton coordinates) needed for posture and activity recognition are acquired using Microsoft Kinect 2.0. Several activities are marked as potentially dangerous and generate alarms sent to caregivers upon detection. Recognition models are developed using Averaged Hidden Markov Models with multiple learning sequences. Action recognition includes methods for dierentiating between normal and potentially dangerous activities (e.g., self-aggressive autistic behavior) using the same motion trajectory. Some activity recognition examples and results are presented.