Novel Dynamic Segmentation for Human-Posture Learning System Using Hidden Logistic Regression

Abstract

In this letter, we propose a novel automatic-segmentation technique for a dynamic human-posture learning system using skeletal-graph time-series. The focused problem is very challenging as no fixed segment-size is appropriate for capturing precise human postures. Our proposed novel dynamic-segmentation scheme will first estimate the number of segments and then the optimal segmentation can be determined using hidden logistic regression subject to the estimated number of segments. Experimental results from the realworld Kinect data are compared with the well-known dynamic-time-warping (DTW) segmentation method. Based on our experiments, our proposed new scheme greatly outperforms the DTW method in terms of miss-detection probability and miss-alignment percentage.