0
The production of large-scale data for understanding and monitoring animal behaviour continues to be a challenge due to high processing time with current automated methods.
Objective: The current study aimed at reducing the processing time with minimal error by (1) developing an efficient approach for applying high performance deep learning methods to feeding and drinking behaviour detection in pigs, (2) validate this approach for use across the entire growing period of pigs, and (3) reduce computing time by utilising instantaneous frame sampling and validating the appropriate sample interval for feeding and drinking behaviour.
Methods: To fulfil this aim, the study included video recordings of seven pens of growing pigs from 30 to 100 kg. The training data included 48 h of video observed across three pens, eight dates (representing eight different weeks) per pen, and two hours per day from 1100 to 1300 h, the validation data included 24 h of video observed across three pens, four dates (weeks) per pen, and same two hours per day, and the full day data included 48 h of video observed for one pen across two full days (day 32 and 53 after insertion). The developed approach relied on a deep learning background algorithm for object and key point detection that for each frame assigned each pig in a pen with a rotated bounding box, the direction of the head and the snout key-point. Using these, simple computer vision calculations were performed to estimate data on feeder and drinker occupation based on intersection-over-union and distance-in-pixels.
Results: This approach performed with errors below 45 s per hour of observation and balanced accuracy above 0.97 for feeding and 0.93 for drinking, with narrow 95% confidence intervals across datasets, indicating high-agreement estimates of feeder and drinker occupation within the tested context. A frame sample evaluation was conducted by testing intervals of 1, 2, 5, 10, 20 and 30 s, and 1, 2, 5, and 10 min. Based on visual inspection of error and bias development with increasing sample interval, a frame sample interval of 10 s was evaluated as most appropriate to retain accuracy and precision while lowering computation power needed.
Conclusion: The developed approach combined with the 10 s sample interval was able to produce large-scale data using 6–7 min for analysing one day of data per pen. The reduced time for computation makes it feasible to use the developed approach to produce large scale data of feeding and drinking in growing pigs.
Larsen MLV, Liu D, Franchi, Tomas Norton GA, Pedersen LJ. Computer vision and ethological methods to produce large-scale data on feeding and drinking behaviour in finishing pigs. Applied Animal Behaviour Science, 2026, 107101, ISSN 0168-1591. https://doi.org/10.1016/j.applanim.2026.107101.