Review of Three Dimensional Human Action Recognition

Three-dimensional human action recognition (HAR) is crucial in many research domains, such as robotics, sports, surveillance and virtual reality. However, to achieve high accuracy and efficiency in performing HAR has always been a challenging task due to variety of pose, appearance and view of the subject. Due to the advanced of processor, the recent state-of-the-art methods to perform HAR are mostly using generic deep learning neural network, such as Recurrent Neural Network, Convolutional Neural Network and Transformer-based Neural Network. Transformer-based Neural Network is highlighted as the most efficient state-of-the-art method to perform action recognition task, with the present of attention mechanism to gradually increase the learning rate of the neural network. A comparative HAR result on three different state-of-the-art methods is tested on three datasets, i.e., MSR-Action3D, NTU RGB+D 60 and NTU RGB+D 120.