Recognition of human-human interaction using CWDTW

Understanding the activities of human is a challenging task in Computer Vision. Identifying the activities of human from videos and predicting their activity class label is the key functionality of Human Activity Recognition system. In general the major issues of Human activity recognition system is to identify the activities with or without the concurrent movement of body parts, occlusion, incremental learning etc. Among these issues the major difficulty lies in detecting the activity of human performing the interactions with or without the concurrent movement of their body parts. This paper aims in resolving the aforementioned problem. Here, the frames are extracted from the videos using the conventional frame extraction techniques. A pixel-based Local Binary Similarity Pattern background subtraction algorithm is used to detect the foreground from the extracted frames. The features are extracted from the detected foreground using Histogram of oriented Gradients and pyramidal feature extraction technique. A 20-point Microsoft human kinematic model is constructed using the set of features present in the frame and supervised temporal-stochastic neighbor embedding is applied to transform a high dimensional data to a low dimensional data. K-means clustering is then applied to produce a bag of key poses. The classifier Constrained Weighted Dynamic Time Warping(CWDTW) is used for the final generation of activity class label. Experimental results show the higher recognition rate achieved for various interactions with the benchmarking datasets such as Kinect Interaction dataset and Gaming dataset.