Multi Activity Recognition Based on Bodymodel-Derived Primitives

We propose a novel model-based approach to activity recognition using high-level primitives that are derived from a human body model estimated from sensor data. Using short but fixed positions of the hands and turning points of hand movements, a continuous data stream is segmented in short segments of interest. Within these segments, joint boosting enables the automatic discovery of important and distinctive features ranging from motion over posture to location. To demonstrate the feasibility of our approach we present the user-dependent and across-user results of a study with 8 participants. The specific scenario that we study is composed of 20 activities in quality inspection of a car production process.