Motion Units: Generalized Sequence Modeling of Human Activities for Sensor-Based Activity Recognition
Hui Liu, Yale Hartmann, Tanja Schultz
Abstract
This paper proposes an innovative activity modeling method for human activity recognition, which partitions the human activity into a sequence of shared, meaningful, and activity distinguishing states, called Motion Units, analog to phonemes in speech recognition. The partitions and generalization define a human activity dictionary, which endows this method with operability, universality, and expandability. Our preliminary experiments demonstrate on-par accuracy with other models while requiring fewer parameters and increasing separability between phases. Furthermore, the developed model was easily transferred with minor adjustments to two other datasets, demonstrating the proposed method's scalability. This framework enables expandable, interpretable, and scaleable modeling and recognition of human activities.