A Dynamic Time Warping Based Kernel for 3D Action Recognition Using Kinect Depth Sensor

Since the advent of 3D sensors and the introduction of robust algorithms for extracting position of skeletal joints out of RGB images, 3D action recognition has attracted a lot of attention. In this paper, we propose a new method for classification of time series representing the position of skeletal joints during the time that an action is performed. One of the main challenges in this classification is the varying speed of performing actions by different individuals. We propose a new kernel function based on Dynamic Time Warping (DTW) for Support Vector Machine (SVM) classification of different length time series. The proposed kernel is applied to each coordinate of every joint and the resulting feature vector is classified using SVM. The effectiveness of proposed method is verified by experiments on two datasets, TST fall detection and UTD-MHAD, for which we achieved the recognition rates of 98.8% and 98.75 % respectively, which are significantly higher than the state-of-the-art performances on these datasets.