Hierarchical Model for Zero-shot Activity Recognition using Wearable Sensors
We present a hierarchical framework for zero-shot human-activity recognition that recognizes unseen activities by the combinations of preliminarily learned basic actions and involved objects. The presented framework consists of gaze-guided object recognition module, myo-armband based action recognition module, and the activity recognition module, which combines results from both action and object module to detect complex activities. Both object and action recognition modules are based on deep neural network. Unlike conventional models, the proposed framework does not need retraining for recognition of an unseen activity, if the activity can be represented by a combination of the predefined basic actions and objects. This framework brings competitive advantage to industry in terms of the service-deployment cost. The experimental results showed that the proposed model could recognize three types of activities with precision of 77% and recall rate of 82%, which is comparable to a baseline method based on supervised learning.