Towards human interaction analysis

Abstract

Modeling and recognizing human behaviors in a visual surveillance task is receiving increasing attention from computer vision and machine learning researchers. Such a system should deal in particularly with detecting when interactions between people occur and classifying the type of interaction. In this work we study a flexible model for detecting human interactions. This has been done by detecting the people in the scene and retrieving their corresponding pose and position sequentially in each frame of the video. To achieve this goal our work relies on robust object detection algorithm which is based on discriminatively trained part based models to detect the human bodies in videos. We apply a ‘Gaussian Mixture Models based’ method for background subtraction and human segmentation. The output from the segmentation method which is labeled human body is combined with the background subtraction methods to obtain a bounding box around each person in images to improve the task of human body pose detection. To gain more precise pose detection models, we trained the algorithm on large, challenging but reliable dataset (PASCAL 2010). Our method is applied in home-made database comprising depth data from Kinect sensors. After successfully getting in every image sequence the corresponding label for each person as well as their pose and position, understanding of human motion comes naturally which is an important step towards human interaction analysis.