Real-time posture and activity recognition

We propose an algorithm for human posture and activity recognition for uncompressed and compressed (MPEG) video inputs. A real-time compression domain technique is developed to recognize different postures such as standing, pointing left/right, opening arms, etc. by using an eigenspace representation of human silhouettes obtained from AC-DCT coefficients. The system stores frames with specific postures and finds the global activity of the human body in the compressed domain. In the uncompressed domain, this information is used as an input for the activity/gesture recognition algorithm. The first part of our approach is invariant to changes in intensity, color and textures and has the advantage of using the available data in the standard compression algorithms. The second part of the system can recognize activities in a set of frames starting with a recognized posture that is classified as a reference movement by the system. A prototype system has been developed with two camera nodes; each consists of a standard camera and a video processing board.