3D cylindrical trace transform based feature extraction for effective human action classification
Georgios Goudelis, Georgios Tsatiris, Kostas Karpouzis, Stefanos Kollias · 2017
Human action recognition is currently one of the hottest areas in pattern recognition and machine intelligence. Its applications vary from console and exertion gaming and human computer interaction to automated surveillance and assistive environments. In this paper, we present a novel feature extraction method for action recognition, extending the capabilities of the Trace transform to the 3D domain. We define the notion of a 3D form of the Trace transform on discrete volumes extracted from spatio-temporal image sequences. On a second level, we propose the combination of the novel transform, named 3D Cylindrical Trace Transform, with Selective Spatio-Temporal Interest Points, in a feature extraction scheme called Volumetric Triple Features, which manages to capture the valuable geometrical distribution of interest points in spatio-temporal sequences and to give prominence to their action-discriminant geometrical correlations. The technique provides noise robust, distortion invariant and temporally sensitive features for the classification of human actions. Experiments on different challenging action recognition datasets provided impressive results indicating the efficiency of the proposed transform and of the overall proposed scheme for the specific task.