A New Representation of Skeleton Sequences for 3D Action Recognition

Skeleton sequences provide 3D trajectories of human skeleton joints. The spatial temporal information is very important for action recognition. Considering that deep convolutional neural network (CNN) is very powerful for feature learning in images, in this paper, we propose to transform a skeleton sequence into an image-based representation for spatial temporal information learning with CNN. Specifically, for each channel of the 3D coordinates, we represent the sequence into a clip with several gray images, which represent multiple spatial structural information of the joints. Those images are fed to a deep CNN to learn high-level features. The CNN features of all the three clips at the same time-step are concatenated in a feature vector. Each feature vector represents the temporal information of the entire skeleton sequence and one particular spatial relationship of the joints. We then propose a Multi-Task Learning Network (MTLN) to jointly process the feature vectors of all time-steps in parallel for action recognition. Experimental results clearly show the effectiveness of the proposed new representation and feature learning method for 3D action recognition.