Particle Filter-Based Predictive Tracking for Robust Fish Counting

In this paper we study the use of computer vision techniques for for underwater visual tracking and counting of fishes in vivo. The methodology is based on the application of a Bayesian filtering technique that enables tracking of objects whose number may vary over time. Unlike existing fish-counting methods, this approach provides adequate means for the acquisition of relevant information about characteristics of different fish species such as swimming ability, time of migration and peak flow rates. The system is also able to estimate fish trajectories over time, which can be further used to study their behaviors when swimming in regions of interest. Our experiments demonstrate that the proposed method can operate reliably under severe environmental changes (e.g. variations in water turbidity) and handle problems such as occlusions or large inter-frame motions. The proposed approach was successfully validated with real-world video streams, achieving overall accuracy as high as 81%.