Deep Learning Based Pervasive Activity Recognition

One of the core problems of pervasive computing is recognizing activities of human beings. Activity recognition under pervasive environments requests implicit of data-collecting devices to users, meanwhile capable of identifying multiple complex activities in daily life. This raises higher requirement for features extracted from data. Recognition accuracy of conventional features and single model is normally unsatisfying. To improve the recognition accuracy, an activity-recognition method under pervasive environments based on deep learning was put forward. Firstly, this method acquires deep features, which are independent of orientation and placement of devices, from acceleration signs through learning, meanwhile obtains vital deep features from vital signs by learning. Then, the ensemble learning approach of multiway stacking is adopted for score level fusion of the above features. Finally, complex activities are recognized based on the acquired features and strategies of score level fusion. A series of experiments demonstrated better performance of the features obtained through learning over conventional features, while the recognition accuracy can be further improved by approach of ensemble learning.