Human action recognition method based on multi-view semi-supervised ensemble learning

Mass labeled data are hard to get in mobile devices. Inadequate training leads to bad performance of classifiers in human action recognition. To tackle this problem, a multi-view semi-supervised ensemble learning method was proposed. First, data of two different inertial sensors was used to construct two feature views. Two feature views and two base classifiers were combined to construct co-training framework. Then, the confidence degree was redefined in multi-class task and was combined with active learning method to control predict pseudo-label result in each iteration. Finally, extended training data was used as input to train LightGBM. Experiments show that the method has good performance in precision rate, recall rate and F1 value, which can effectively detect different human action.