情報処理学会 第77回全国大会講演要旨

This paper proposes a method to track 3D motion data of a fish group in a fish tank using multiple cameras outside of the tank. As fish are almost identical, occlusion among individuals and reflection at the fish-tank surface make it difficult to maintain the object association during the tracking. To tackle this difficulty, we combine a 3D fish shape model consisting of a flexible midline and fixed elliptical cross sections with the mixture particle filter to exploit both shape and motion priors. During the calculation of the likelihood function, we utilize a pixel-wise varifocal camera model to achieve an efficient forward projection from underwater 3D object to each of image planes considering the effect of light refraction.