Fuzzy centralized coordinate learning and hybrid loss for human activity recognition

Abstract

Human activity recognition has been a popular research topic in recent years. The rapid development of deep learning techniques has greatly helped researchers to achieve success in this field. But the researchers usually over look the distribution of features in the coordinate space despite its significant effect on the convergence status of network and classification of activities. This paper proposes a combined method based on fuzzy centralized coordinate learning (FCCL) and a hybrid loss function to overcome the explained constraint. The FCCL induces features to be dispersedly spanned across all quadrants of the coordinate space. For this reason, the angle between the feature vectors of the activity classes increases significantly. Furthermore, a hybrid loss function is presented to increase the discriminative power of the proposed method. Our experiments were carried out on the opportunity and the PAMAP2 datasets. The proposed method has been compared with six machine learning and three deep learning methods for activity recognition. Experimental results showed that the proposed method outperformed all of the comparative methods due to identifying discriminative features. The proposed method successfully enhanced the average accuracy by 17.01% and 3.96% on the PAMAP2 and opportunity datasets, respectively, compared to the deep learning methods.