3D human activity recognition system using motion history volume

Abstract

Activity recognition is an important and challenging topic in computer vision, with many interesting applications including interactive environments, visual surveillance, advanced user interface, personalized sports training, choreography, and clinical research of orthopedic patients. However, most works show view-specific recognition so recently researches have begun to provide view-invariant human activity recognition in those applications. One approach is to use Motion History Volume (MHV) and apply Fourier transformation to the MHV to have the descriptor view-invariant characteristic. In this paper, I describe a 3D human activity recognition system that are based on the approach and can recognize ten activities: walking, running, sitting down, standing up, falling down, punching, kicking, turning, hugging, shacking, from sequence of images from multiple cameras installed in a lobby or a corridor in the interior of a building. I made an experiment on seven subjects. Although they conducted ten activities in a slightly different manner, the system successfully recognized ten human activities.