Weight Estimation of Broilers in Images Using 3D Prior Knowledge

摘要

Cameras are already widely used for inspection and monitoring tasks in poultry slaughter houses. In this paper we evaluate the use of computer vision for broiler carcass weight estimation. We compare the use of 2D image features with 3D features extracted from a statistical shape model fitted to the image. The statistical shape model is built from 45 3D scans captured from broiler carcasses collected at a slaughter house. The use of this 3D prior gave a reduction in mean absolute error compared to 2D features alone and achieved an overall mean average percentage error of 3.47%. The algorithm can run real time and was tested on a dataset containing 136,472 images of broilers, captured at a real production site.