Human Action Recognition Based on Angle Descriptor

A simple and effective method for 3D skeleton based action recognition is proposed in this paper. Instead of taking the whole skeleton joints as the input, we select several active joints to represent the entire action which motion ranges are relatively large via evaluating their variance and give them different weights. Then by calculating the angles between these joints and the center joint in their three projections produces a feature set at each frame which is applied in a bag-of-words to form the 2D array. The final features are cascaded by these 2D arrays. During this process, the feature numbers can be reduced effectively. The random forest is utilized to classify different actions. Experiments on MSR-Action3D dataset demonstrate that our approach is able to achieve the state-of-the-art performance with high recognition rate and computational efficiency.