An efficient skeleton-based action recognition approach with view transformation

Recently, human action recognition has made satisfactory progress on almost daily activities. The emergence of depth sensors makes it more convenient to obtain human skeleton information from videos. Many problems in action recognition can be solved by using joint points, such as light changes and background noise. However, the unfixed viewpoint is one of the difficulties that affect action recognition, which is still under-explored for most existing skeleton-based works. Besides, training a complex deep model increases hardware complexity and cost. The proposed model uses a view adaptive asymmetric convolutional network (VA-ACN) to extract skeleton features from the skeleton data and transform the original body perspective into a new observable viewpoint. The ResNet50-based backbone is improved for high-performance classification. Compared with recent state-of-the-art models, the proposed model improves feature extraction and classification performance without increasing extra computation cost. Hence, it saves training time while improving the performance of action recognition under certain conditions. Experimental results show that the proposed model outperforms state-of-the-art methods.