Towards an electronic system for estrus detection in swine

Because most pigs worldwide are now bred by artificial insemination (AI), accurate estrus detection is essential. There is consensus that display of the lordosis response (“standing”) following physical boar contact or application of the back-pressure test (BPT) is the gold standard for estrus detection in pigs. However, little research has explored its true accuracy and evaluated whether automated measures could improve upon such a manually driven process. On commercial pig breeding farms, estrus detection relies on multiple trained technicians and daily efforts and occurs year-round. This places an undue burden on an already volatile and limited farm labor supply. And while research into automated technology specific to swine estrus detection is evolving, it remains underdeveloped. Although electronic estrus detection (EED) systems have been successfully integrated into the management of dairy and goat herds, use in swine breeding herds has faced challenges. Therefore, the aim of this thesis is to summarize current knowledge of estrus detection procedures in swine and extrapolate how these processes can be improved using computer vision technology. A YOLOv8 object detection model for distinguishing erect ear posture from neutral ear posture had a mean average precision (mAP) of 98.3% at 0.5 intersection over union (IoU) for all classes. This is acceptable for automatic detection of ear posture in gilts. A YOLOv8 instance segmentation model detected the outline of gilts in single-stall housing with an overall mAP of 99.5% at 0.5 IoU; this model's performance was also satisfactory. A YOLOv8 movement classification model assigning three classes of behavioral responses to gilts receiving the back-pressure test (BPT) showed high false positive rates for one class and high false negative rates for two classes. This model is not yet satisfactory in automatically classifying types of behavioral responses to the BPT. Additionally, a YOLOv8 object detection model was trained to identify four postures (sitting, standing, kneeling, and lying) in prepubertal gilts, achieving an overall average of 0.976 mAP at 0.5 IoU. This model was then applied to a 24-hour dataset of gilts transitioning from proestrus into estrus and metestrus. Time budgets were created for each gilt according to the four postures detected. No statistically significant differences existed among the four postures between pre-puberty and the conclusion of estrus.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Animal Sciences

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2023

Author dc:creator

• Bushman, Jeni L.

Contributors dc:contributor

• Condotta, Isabella
• Green-Miller, Angela
• Knox, Robert

• Machine Learning
• Computer Vision
• Estrus Detection
• Back-pressure Test

Statement dc:rights

• Copyright 2023 Jeni Bushman

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/122042

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F122042