Model-based detection of pigs in images under sub-optimal conditions

Abstract The automatic detection of pigs in camera images from within the barn helps scientists and farmers to detect abnormal behaviour or problematic housing conditions and to investigate the causes. An established method for determining the position of pigs is the binary segmentation of the image and the subsequent modeling of the individual animals. Many studies are based on elliptical models because they sufficiently reproduce the positions of the pigs with a few parameters. However, the existing methods for adapting the ellipses require an almost perfect segmentation as they depend on the clear delimitation of individual animals. Although the animals are usually visually distinct from the background, a uniform segmentation is not always feasible. Due to occlusions, dirt or shadows in the barn, incomplete or faulty segmentation can occur even with advanced segmentation techniques. So this paper introduces a novel method for adapting the ellipses, which is not based on the edges of the segmentation but looks at all segmented pixels. This makes it easier to compensate minor errors in segmentation and helps to process images even under sub-optimal conditions, such as poor lighting or unfavourable camera positioning.