Joint movement similarities for robust 3D action recognition using skeletal data

• Action recognition using a combination of spatio-temporal skeleton based features.
• New data-driven methods for extracting discriminative features in each action class.
• A new similarity function based on LCSS for classification of new action instances.
• Evaluation of the approach using Kinect and mocap data showed improved performance.
• The approach achieved higher classification rate for noisy data in contrast to DTW.

Human action analysis based on 3D imaging is an emerging topic. This paper presents an approach for the problem of action recognition using information from a number of action descriptors calculated from a skeleton fitted to the body of a tracked subject. In the proposed approach, a novel technique that automatically determines discriminative sequences of relative joint positions for each action class is employed. In addition, we use an extended formulation of the longest common subsequence algorithm as a similarity function, which allows the classifier to reliably find the best match for extracted features from noisy skeletal data. The proposed approach is evaluated using two existing datasets from the literature, one captured using a Microsoft Kinect camera and the other using a motion capture system. The experimental results show that the approach outperforms existing skeleton-based algorithms in terms of its classification accuracy and is more robust in the presence of noise when compared to the dynamic time warping algorithm for human action recognition.