Toward Recognition of Short and Non-repetitive Activities from Wearable Sensors

Activity recognition has gained a lot of interest in recent years due to its potential and usefulness for context-aware computing. Most approaches for activity recognition focus on repetitive or long time patterns within the data. There is however high interest in recognizing very short activities as well, such as pushing and pulling an oil stick or opening an oil container as sub-tasks of checking the oil level in a car. This paper presents a method for the latter type of activity recognition using start and end postures (short fixed positions of the wrist) in order to identify segments of interest in a continuous data stream. Experiments show high discriminative power for using postures to recognize short activities in continuous recordings. Additionally, classifications using postures and HMMs for recognition are combined.