Environmentally adaptive fish or no-fish classification for river video fish counters using high-performance desktop and embedded hardware
Jürgen Soom, Vishwajeet Pattanaik, Mairo Leier, Jeffrey A. Tuhtan
Abstract
Automated fish counters featuring robust, real-time computer vision capabilities can provide a cost-effective means to count migrating freshwater fish. In this work, we propose a four-stage process for automatically sorting videos with and without fish. Underwater fish counter videos provide a challenging range of environmental conditions including clear water, biofilm growth, bubbles, turbidity, low light and overexposure. To address this, our method also includes the automated classification of these six environmental conditions. The proposed methods are computationally efficient and can be implemented on servers, high-performance desktop computers and low-cost, energy-efficient embedded hardware. The models were trained, tested, and validated using a collection of 3000 videos taken from underwater fish counter installations in several alpine and lowland European rivers provided by commercial and governmental collaborators. This work demonstrates a fast, accurate, and robust computer vision workflow for large-scale automated freshwater fish counting systems.