Recently, various motion measurement systems are used as a basic technique for motion analysis. These systems include electrogoniometric method and optical method by land markers put on the body surface. However, the former methods restrict human motion, and the land markers are missed and dislocated by deformation of the body surface. Additionally in clinical application, more simple method is required, which can reduce troubles with walking and fitting up the markers. In this study, we developed a new technique of motion measurement to fit the human geometrical model into the video image, in order to reduce the difficulties of measurement by land markers. The geometrical model contained 16 segments. Each segment had 1 degree of freedom of rotation and 2 degrees of freedom of translation. Human motion was recorded to a LASER disc recorder (TEAC: LA200A). Then the video image was played frame by frame and sent to a personal computer (TOSHIBA: J 3300/50) using a frame memory (HAMAMATSU: C1900). Subsequently, the body image was picked up using the thresholding method in image processing, and the geometrical model was fitted into this image. As the first step, the model was corrected roughly using a distance-transformed image which corresponded nearly to the center line of the body image. After rough fitting, the segment of the model was divided into 4 or 6 parts, and the model was corrected finely based on the direction and the value determined from the overlapping area between those divided parts and the body image. Repeating the fine correction, the joint angles can be estimated to about 1 degree. Using this method, we measured the normal gait motion with a CCD camera set at the distance of 3 meters. It takes about 20 seconds per frame to calculate, or about 10 minutes per 1 cycle of walking (30 frames). The amplitude of joint angle measured by this method agreed with what was measured by land markers simultaneously; the angles calculated by land markers had an offset because of the dislocation of the markers put on the joint. There was a tendency to decrease amplitude of shoulder joint angle with land markers and their cables. The influence of noise of image and model size error within 5% of its segment length are almost neglected in calculation of joint angles. This is because this method is using the information of overlapped area for fitting models. There still remain several problems in this method, such as automatic determination of the model size and detection from normal scene which includes background, clothes and tools. However, the developed method is characterized by unrestricted motions and simplified measurement.