AQUA 2024

August 26 - 30, 2024

Copenhagen, Denmark

FISH BEHAVIOR IDENTIFICATION BASED ON COMPUTER VISION

H . Alvheim* ,  S.M. Jakobsen,  E. Kelasidi and M. Føre

Department of Engineering Cybernetics , Trondheim, Norway,

Department of Aquaculture Engineering, SINTEF Ocean, Trondheim, Norway

*E-mail : hannegrete.alvheim@gmail.com

 



 The aquaculture industry needs to address several challenges to secure the demand for increased seafood production and sustainability. When targeting increased autonomy and more exposed locations within the fish farming industry [1] , and  when introducing intrusive objects , sensors and robots into these environments care should be taken to ensure good fish welfare . This paper addresses this by developing, implementing, and comparing methods to identify fish behavior when affected by intrusive objects. A novel  approach for detecting , tracking, and estimating the 3D position of individual fish has been created to reach this goal. T he  proposed  method and its  versions  utilize a combination of stereo vision by SuperGlue , triangulation and RAFT-Stereo, image pre-processing including morphological area opening and discrete wavelet transform (MO-WT), object detection by YOLOv8,  and multi-object tracking by ByteTrack to estimate the relative distance between a stereo camera system and detected fishtails.  Additionally, these methods can by Savitzky -Golay smoothing  of raw estimates  derive indirect behavioral estimates of fish, as shown in Figur e 1.

The methods have been tested  on d ata collected from  an industrial fish  farm  with Atlantic Salmon  on the SINTEF ACE site [2] in 2023. A stereo camera system and two sonars were put on a  structure coated with different color- and shape combinations ( yellow, white,  cube, big cylinder and small cylinder)  placed at 8 m depth. Six replicate experiments of  12 min were conducted for each color- structure combination.

Results are shown in Tabl e 1 , concluding  that fish stay further away from yellow and large objects than from white and small objects. These conclusions are aligned with those found when processing sonar data  from the same cage [3] . Additionally, the system can estimate derived parameters close to real-time as well as over longer time periods .  The proposed method is general and can  thus  be used to study both the collective and individual behavior of fish.  This advantage  makes the method especially useful, both for instant feedback in control systems and for observing altered fish behavior over time, and it can be  an essential contribution in the future of fish farming. 

References

 [1] E. Kelasidi and E. Svendsen, Robotics for Sea-Based Fish Farming.  Springer Int. Publishing, 2023, pp. 1–20

[2] “Ace,” https://www.sintef.no/en/all-laboratories/ace/, accessed: 2023-12-19

[3 ] Q. Zhang, et. al,, “Farmed atlantic salmon (salmo salar l.) avoid intrusive objects in cages: The influence of object shape, size and colour, and fish length,” Aquaculture, p. 740429, 2023