Multiple-Stage Classification of Human Poses while Watching Television

We compared the accuracy measure between a single-stage classifier model and a multiple-stage classifier model in postural classifications using Kinect. Postural training sets were collected from Kinect's skeletal data streams, based on some of the common human postures during television watching. Three types of training sets were used, including Kinect's raw skeletal training set, skeletons with attribute selection training set, and skeletal position transformation training set. We selected four learning models, namely, neural network, naïve Bayes, logistic regression, and decision tree, for learning our data sets and classifying a testing set to find the appropriate learning model. The best accuracy value of our experiment was 87.68 % by using skeletal position transformation training set with neural network. In the future, we will apply our technique and methodology to track elderly behaviors while they are watching television.