Automatic live fingerlings counting using computer vision

Fish counting is still a rudimentary process in most fisheries in Brazil. Current solutions are generally unaffordable for small and medium-size producers; hence, in order to provide a low-cost solution, this paper proposes a new technique for fish counting and presents a new image dataset to evaluate fish counting systems. The dataset is composed of a series of videos partially annotated at frame-level, which include approximately a thousand fish in high-resolution images. We describe a computer-vision based system that counts fish by combining information from blob detection, mixture of Gaussians and a Kalman filter. This work shows that the proposed method is a feasible approach for automatic fish counting, reducing costs and boosting production, as it increases labor availability. Our approach is efficient for fingerlings counting, with an average precision of 97.47%, recall of 97.61% and F-measure of 97.52% in the provided dataset.