Abstract Automated behavior analytics can make continuous welfare and productivity monitoring feasible on a commercial scale. However, independent field validation against human observers in commercial-type settings remains limited. The objective of this study was to validate a commercial AI-driven, computer vision behavioral analytics platform (PigGuard, Serket. Amsterdam, Netherlands) by comparison with manual human classification (ground truth) and assess performance across various pig ages. A 13-week growing-finishing pig study was conducted in 2 rooms at the North Carolina State University swine nutrition barn. Each room contained 12 pens, with seven pigs placed in each pen at approximately 10 weeks of age. Rooms followed a weekly schedule that randomly cycled 3 thermal conditions (thermoneutral, moderated heat, and moderated cold) for 2 consecutive days each. Continuous overhead Red-Green-Blue RGB video surveillance was collected for each pen at 20 fps. For the behavioral validation subset, week 1 (early finishing), week 6 (mid), and week 12 (late) were selected. One room was analyzed per week of interest. For every week and each thermal condition, a single 30-minute segment was extracted at 2:00, 8:00, 14:00, and 20:00 h. As each condition was applied for 2 consecutive days, the sampled day for each segment was chosen at random, except for the 08:00 segment, which was always selected from day 2 to ensure consistent thermal conditions were applied. Trained human observers used BORIS (Behavioral Observation Research Interactive Software) to annotate videos using a predefined ethogram, capturing 4 main behaviors: active, inactive, feeding, and drinking. Human labels will be compared to software outputs from PigGuard. Analysis is ongoing. Results will be compared using precision, recall, and F1 scoring. Insights will inform accuracy of the computer vision program across grow-finish pig age and thermal conditions. Computer vision systems such as PigGuard are a promising tool for commercial settings; this validation will quantify PigGuard’s level of agreement with human annotations and demonstrate its potential capabilities and limitations as an innovative swine management tool.
Suazo et al. (Wed,) studied this question.