Author(s): Aleksandr Ivanov; Alexandra Kolosova; Jeffrey Andrew Tuhtan; Bernd Mockenhaupt; Nicole Scheifhacken

Linked Author(s): Jeffrey Tuhtan

Keywords: Fish length estimation; Monocular camera; Video tunnel; Deep learning; Computer vision

Abstract: Fish body length is an essential ecological parameter for assessing population structure, growth, and migration patterns in river systems. Here, we investigate the feasibility of automated fish body length estimation in fishpasses using video tunnels with a single RGB camera system and deep learning. The proposed approach integrates YOLOv5s for fish detection with a ResNet18-based feature extractor and linear regression layers for length estimation. The model was trained on a dataset of 4,111 videos recorded in the Main, Elde, and Mosel rivers in Germany. Evaluation on 280 previously unseen test videos covering fish lengths from 10 to 75 cm achieved an overall F1-score of 0.84, increasing to 0.92 for fish up to 60 cm. The results demonstrate that, despite certain limitations associated with local environmental conditions and partial visibility of larger fish, monocular length estimation can be a viable alternative when upgrading to stereo-vision systems is not feasible.

DOI: https://doi.org/10.64697/iahr.proc.ise2026.abs.123

Year: 2026