Vision-based measurement of fast ball movement for sports simulation

Abstract

In this thesis, we present an algorithm to measure precise movement of fast ball from multi-exposed images. Recently, several computer-simulated ball sports games have been developed and some of them such as golf and soccer have already achieved commercial success in market. Most sensors for the interactive simulation, however, are too expensive or have limitations on its physical sensing ability and spatial sensing coverage. A multi-camera system with a stroboscope is a suitable sensor system for interactive sports simulation, which provides accurate sensing with spatial freedom at low-cost. Nevertheless, it also has several problems since it uses a stroboscope light source and does multi-exposure for each camera frame. For instance, strong top-down illumination makes fade on the ball, and following club head overlaps the ball image. In this thesis, we addressed several challenges arisen while we measure the movement of fast ball with this system, and presented our solution. We have proved that our algorithm could get precise location of the shaded ball. We have tested it with various ball sizes at various angle, and our algorithm could obtain the center of the ball in sub-pixel resolution. In addition, the issue that ball images could be corrupted by other objects in multi-exposure was solved by detecting outliers, and interpolating missing balls. We have shown that until 33% of balls were missing, our algorithm could maintain its stability, and in cases that more balls were missing, it still returned correct result with high ratio. We also showed that the performance of multi-camera single stroboscope sensor system with our algorithm was very close to that of a high performance radar sensor system, TrackMan. Finally, we showed that the proposed algorithm ran within 50 milliseconds, which is enough for real-time sports simulation.