摘要：Detecting human interactions in a public place where no physical touch occurs has important applications in many surveillance tasks. In this paper, we explore the possibilities to automatically detect such distant human interactions without recognizing the specific human actions. Specifically, we use a highly simplified formulation of the interaction in this paper: 1) when a person does not interact with others, he always performs a non-interactive action that is largely periodical, such as walking, 2) when two persons interact with each other, they both perform a same short-duration interactive action, such as waving hand, that are different from their non-interactive actions. Based on this formulation, we develop a new approach to localize the subvideos that describes the interactive actions from Kinect videos, which provide both the RGB and depth information. We then develop a new approach to compare the pose and kinematic features in these subvideos (from different people) to see whether they describe a same interactive action. Without any supervised learning and action recognition, the proposed approaches are not limited to a specified set of interactive and non-interactive actions. In the experiments, we justify the performance of the proposed approaches on 100 Kinect videos with 10 different interactive actions.
关键词：distant human interaction;abnormal detect;dynamic time warping
