Combined deep centralized coordinate learning and hybrid loss for human activity recognition

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. During the training process with deep learning techniques, features and time dependencies between them are well learned. However, researchers generally ignore the distribution of extracted features in the coordinate space despite their significant effect on classification and network convergence status. The present article utilizes a simple but effective centralized coordinate learning method that dispersedly spans extracted features across the coordinate space. This method causes the angle between the features of different classes to increase significantly. A hybrid loss function is also suggested to enhance the discriminative power of learned features. Some experiments were carried out on the OPPORTUNITY and the PAMAP2 datasets. The results showed that the proposed method outperformed the recently proposed deep learning methods, including the Deep ConvLSTM, CNN-LSTM-ELM, and Hybrid methods. This high efficiency was due to the identification of discriminative features.