Stochastic multiple fish tracking using motion and shape consistency

Conventional appearance-based multiple target tracking methods could fail in handling occlusions in fish surveillance video, since the intensities of fishes are similar with each other. In view of this challenge, this paper proposes to exploit both the motion and the shape feature to differentiate multiple fishes, and incorporate the motion consistency and the shape consistency into a Bayesian inference framework to find the maximizing a posterior (MAP) estimations of fish contours and labels. Experimental results are presented to demonstrate the superior performance of the proposed approach.