Activity recognition is an essential task that enables analysis of human behavior. We focused on recognition of activities from sensors that provide 3D coordinates (such as Vicon and Ubisense) using machine learning algorithms. We analyzed how sensor placement and amount of noise affects recognition accuracy and developed a series of filters to reduce noise in both pre- and post-processing phase. These include a filter that enforces human body constraints and Kalman filter in the pre-processing, and a filter for spurious activity transitions, which eliminates infeasible activity transitions, e.g., from lying to standing without standing up in between.

Relevant papers:

• M. Luštrek and B. Kaluža. Fall Detection and Activity Recognition with Machine Learning. Informatica, 33(2):197-204, 2009. [pdf]
• B. Kaluža, V. Mirchevska, E. Dovgan, M. Luštrek, and M. Gams. An Agent-based Approach to Care in Independent Living. Lecture Notes in Computer Science, vol. 6439, pp. 177-186, AmI’10, Malaga, Spain, November, 2010. [pdf]
• M. Luštrek, B. Kaluža, E. Dovgan, B. Pogorelc, and M. Gams. Behavior Analysis Based on Coordinates of Body Tags. Lecture Notes in Computer Science, vol. 5859, pp. 14-23, AmI09, Salzburg, Austria, 2009. [pdf]