Computer vision based monitoring of fattening pigs: how well do pigs behave?

Until now, monitoring of fattening pigs has been focused primarily on production parameters such as feed and
water intake and weight. However, camera monitoring using computer vision based approaches can offer new
opportunities to continuously monitor both production parameters and animal behaviour. Specifically, monitoring
the animal behaviour can enable insight into animal welfare issues such as tail biting, improve the knowledge of
fattening pigs social structure in a pen or can even assist in modelling pig growth. The goal of this research is to
combine traditional PLF technologies (RFID antennas, flow meters, weight scales) with state-of-the-art computer
vision models in order to create the link between growth and behaviour. In this study, 120 pigs (1 compartment, 8
pens) were continuously monitored by an angled top view camera (1 camera per pen). The resulting videos were
analyzed using modular computer vision deep learning models for detection (M3det), tracking (M3track), pose
estimation (M3PoseSeg) and pig interactive behaviour (M3ethology). Furthermore, all pens were equipped with
RFID antennas at the feeder and drinker, water flow was measured at all drinkers, feed intake was measured in
4 pens and pig weight was recorded in each pen by a weight scale at the feeder or drinker. The computer vision
tracking was also assisted by detection of large ear tags and by the RFID identifications to improve the length of
individual pig tracks. Possible features (such as pig activity index) as a result of the computer vision models were
explored and combined with the measured production parameters.