An extremely randomized trees approach for gesture unit segmentation

Automated human gestures analysis has a wide range of promising applications in many advanced fields including human-computer interaction, motion analysis, and security surveillance. However, automatic gesture segmentation is still a very challenging task due to the spatio-temporal variation and endpoint localization issues, and the variation of gestures based on performers, topics and performance sessions. This paper presents a novel framework for segmenting gesture unit based on Ada-Boost and extremely randomized trees algorithms from video streams. In this approach, an Ada-Boost feature selection algorithm is applied to select compact feature subsets from the numerous raw extracted features to reduce the computational time as well as to improve the segmentation rate of the gesture segmentation model; then, selected features are fed to a robust extremely randomized trees classifier, given their capability to handle complex and unbalanced data, to segment gesture unit. The evaluation results of the experiments conducted on the publicly available benchmark gesture segmentation data set indicate that the proposed technique improve the segmentation metric by as much as 5.2% over the previously applied techniques.