Activity recognition using Eigen-joints based on HMM

In this paper, we present an approach for activity recognition by using 3D skeleton data obtained with a Kinect sensor. Primarily, we use the simplified dynamic time wrapping (DTW) and calculate Euclidean geometry distance to obtain the probable activities from the trained data. Afterwards, for each activity, we define a modified activity feature descriptor using the interrelation of correlated joints in each frame. Before classification, we employ normalization to avoid non-uniformity in coordinates, and then Principal Component Analysis (PCA) is applied to deduce redundancy and decrease the dimensionality. As the result Eigen-joints for each activity are obtained. Finally we classify the joints into multiple actions using Hidden Markov Model (HMM). The experimental result on benchmark dataset shows that the accuracy approximates that of the state-of-the-art.