AN INTEGRATED METHOD TO OBTAIN THREE-DIMENSIONAL COORDINATES USING PANNING AND TILTING VIDEO CAMERAS

Kinematic measurements of sports performances have generally been gathered using stationary cameras. Many activities, however, cover dimensions that are much larger than the athlete, resulting in measures that are less accurate than desired. One solution is to increase the resolution of the acquisition equipment. Hardware providing sufficient resolution, if possible to find, may be prohibitively expensive. Another solution is to use rotating cameras to follow the athlete’s movements. Until now, this procedure demanded time-consuming calibrations and expensive, specialized equipment. An efficient, accurate, and inexpensive method was developed to quantify threedimensional motion from rotating video cameras. The procedure uses two 20,000 count/revolution optical encoders embedded in specially machined tripod heads to sense the angular positions of the cameras. One encoder is aligned vertically to measure pan positions, while the other is aligned horizontally to measure tilt. The pan and tilt angles are printed on each video image by an interface unit. The images are then recorded on videotape and imported into a computer using a commercial frame grabber. Locations on the body may be identified manually or automatically if markers are placed on the athlete. As the points are tracked, the pan and tilt angles are decoded and stored with the digitized data. Custom software modules use these data to compute the 3D coordinates via ray-tracing techniques. Calibration, conducted before or after the performance, requires three to nine vertical rods of known length placed near the volume of interest. No on-site measurements or background markers are needed. Furthermore, calibration does not restrict camera positioning; the locations depend upon the focal length of the camera lenses, the resolution of the encoders, the shape of the movement volume, and the speed of the athlete. Dynamic accuracy was measured by moving a 0.900 m long rod through a 2 m x 4 m x 15 m volume while sampling from two 60 Hz video cameras. Nine hundred images were analyzed. The endpoints of the rod were digitized and its length computed for each video image. The mean length of the rod was 0.891 m and the Root Mean Square Error was 0.005 m. These measures compared better than other reported methods. Applications have ranged from multi-stride running in the laboratory or on a track, long jumping, high jumping, gymnastics, pole vaulting, and ski jumping