Efficient Interaction Recognition Through Positive Action Representation
Tao Hu,Xinyan Zhu,Wei Guo,Kehua Su
DOI: https://doi.org/10.1155/2013/795360
IF: 1.43
2013-01-01
Mathematical Problems in Engineering
Abstract:This paper proposes a novel approach to decompose two-person interaction into a Positive Action and a Negative Action for more efficient behavior recognition. A Positive Action plays the decisive role in a two-person exchange. Thus, interaction recognition can be simplified to Positive Action-based recognition, focusing on an action representation of just one person. Recently, a new depth sensor has become widely available, the Microsoft Kinect camera, which provides RGB-D data with 3D spatial information for quantitative analysis. However, there are few publicly accessible test datasets using this camera, to assess two-person interaction recognition approaches. Therefore, we created a new dataset with six types of complex human interactions (i.e., named K3HI), including kicking, pointing, punching, pushing, exchanging an object, and shaking hands. Three types of features were extracted for each Positive Action: joint, plane, and velocity features. We used continuous Hidden Markov Models (HMMs) to evaluate the Positive Action-based interaction recognition method and the traditional two-person interaction recognition approach with our test dataset. Experimental results showed that the proposed recognition technique is more accurate than the traditional method, shortens the sample training time, and therefore achieves comprehensive superiority.