Abstract
We developed a system for automatically monitoring sheep behavior using computer vision. We designed an algorithm for identifying sheep within images. This algorithm applied thresholding to a density histogram generated from mobile objects identified by a frame difference method. To assess the performance of the prototype software, an experiment was conducted using Suffolk sheep. In the experiment, an outdoor fenced area of about 5m x 5m was constructed. Images of the behavior of the sheep were captured from a height of approximately 5m at one-second intervals from sunrise to sunset. The images were taken on both fine and cloudy days. From the captured images, sheep position coordinates were automatically measured and recorded with the prototype software. To examine the effect of the changing solar illumination, position error was calculated for each image sampled at five-minute intervals. A position error of less than 1/4 the width of the sheep occurred at a rate of 75.6% on fine days and 95.2% on cloudy days. To examine the effect of rapid changes in brightness due to cloud movement, we calculated the position error at one-second intervals within 20 seconds of the appearance or disappearance of shadows. Under these conditions, position error less than 1/4 the width of the sheep occurred at a rate of 73.7% when the sheep passed near the water tub and the feeding trough. In other cases it occurred at a rate of 90% or more.
