An image synthesizing method based on human motion recognition from stereo images

An image synthesizing method for the conference system with realistic sensations is discribed. In this conference system, a human image in virtual space has to be synthesized by tracing the motion of human in real space.But it is difficult to detect positions of all human joints from real image. We proposed a new image synthesizing method using an associative memory. This is memorized human representative motion patterns to conpensate insufficiency of detected information from real image.The human body region can easily be detected from each images of stereo by using deffrences between input image and background image, and also skin color regions can easily be detected by color information. 3D-positions of few points of human body(head, face, hands) are obtainable by stereo matching of a representative point in detected regions. By an associative memory memorized relation of standerd motion patterns with parameter that is independent from a human direction and an initial horizontal position, a motion is recognized from time-sequential 3D-positions of few points of human body. After motion recognition, a human direction is detected by matching the head movement of recognition result s and the movement of detected head position. At last 3D-positions of parts of representative motion in virtual space are adjusted to detected 3D-positions of parts by using Jacobian matrix.Simulation results are presented to show an effectiveness of this method.