Gesture Recognition on Kinect Time Series Data Using Dynamic Time Warping and Hidden Markov Models

Abstract

In this paper we analyse the variation of the gesture recognition accuracy of several time series classifiers, based on input provided by two different sensors: Kinect for XBox 360 (Kinect 1) and its improved, newer version, Kinect for XBox One (Kinect 2). This work builds upon a previous study analysing classifiers' performance on pose recognition, considering multiple factors, such as the machine learning methods applied, the sensors used for data collection, as well as data interpretation and sample size. As for the classification of time series gestures, we analyse similar factors, by constructing several one-hand gesture databases that are used to train and test the Dynamic Time Warping (DTW) and Hidden Markov Models (HMM) algorithms. We observed no significant difference in classification accuracy between the results obtained with the two sensors on time series data, although Kinect 2 performs better in pose recognition. Overall, DTW obtained the best accuracy for Kinect 1 time series data, on datasets with fewer samples per class (about 15), the accuracy decreasing drastically with the increase of the number of samples for each class (from 97.8% drops to 66.6%). However, for HMM the accuracy is similar or higher (between 90.7% and 94.9%) for databases with more samples per class (up to 90 entries) than for those with fewer, which makes it preferable to use in a dynamic system.