Human action recognition is one of the most challenging and attractive areas in the field of computer vision. Conventional research on human action recognition has mainly focused on data modality of video or optical flow. However, the human skeletal feature has much stronger expressive power of motion dynamics, which is not sensitive to illumination and scene variation. Owing to the advantages of deep learning approaches on skeleton data in recent years, many pilot approaches have been proposed, which are merited by their significant performance enhancements on both baseline and large-scale datasets. This research investigates these models and their breakthroughs, especially focusing on the graph convolution network (GCN) and skeleton-based data techniques. The report work mainly covers the following aspects: comparing RNN, CNN and GCN-based approaches from the perspective of their operational logics; a detailed review of the best referred models in recent years; a development framework of skeletal feature-based human action recognition framework is proposed with preliminary assessments using benchmarking datasets; and finally, the envisaged future directions for skeletal feature-based human action recognition study are discussed.