Automated visual auditing of swine movement
Hang, Chee Hin
Is Version Of
Agricultural and Biosystems Engineering
This thesis investigated the automated visual auditing of pigs during movement which resulted in a new computational framework that yielded reliable audit results and formative observations for automated animal production systems. Current audit errors, conservatively estimated to be at 5-percent per group movement of animals, have traditionally been dealt with by either increasing the number of human auditors or absorbing the costs of uncertainty into general operating costs. These uncertainty costs can range from a few hundred dollars at the load level to millions of dollars per day when scaled across the United States. The computational approach developed in this work demonstrated an automated visual auditing algorithm that performs at no worse than 2-percent error. Further, the algorithm was shown to significantly outperform human auditors in both accuracy and precision. Formative observations made during this research suggested: i) all swine production facilities move animals through doorways and this provides a common point of visual observation, ii) rushed and unsteady herding of pigs can cause bidirectional movement and clustering occlusions, and iii) manual counts performed by humans are inconsistent and errors increase as the number of pigs audited increases. This research is of importance to animal production researchers and practitioners, as well as those in the fields of robotics and computer vision systems, because it identified a common camera placement for production facility interoperability and there are animal movement factors that can affect both automated and human counting accuracy.