Evaluation of turkey behavior under different night lighting treatments using machine learning

A video recording system was used to record activity of turkeys living in sixteen different pens continuously for six weeks for an animal welfare study assessing differences in behavior for birds raised with night lights compared to birds raised in complete darkness at night. Human scan sampling was performed to count the number of birds eating, resting, and standing once every hour for eight of the pens on day 41 of the study. Machine learning models were trained to detect these simple behaviors, and output from each model was compared to the human observations. This paper discusses preliminary results from the different machine learning object detection models tested, model training challenges, and comparisons between model detections and human observations.