The sorting hat: An automated activity index based on finishing pig behavior

Felton, Mekali

Permalink

https://hdl.handle.net/2142/129666 Copy

Description

Title

The sorting hat: An automated activity index based on finishing pig behavior

Author(s)

Felton, Mekali

Issue Date

2025-04-03

Director of Research (if dissertation) or Advisor (if thesis)

Green-Miller, Angela

Committee Member(s)

• Condotta, Isabella
• Malvandi, Amir

Department of Study

Agricultural & Biological Engr

Discipline

Technical Systems Management

Degree Granting Institution

University of Illinois Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• machine learning
• behavior analysis
• finishing pigs
• yolov8
• mobilenet

Abstract

A new tool is needed to support animal behavior observation on pig production farms. New guidelines for pig production include behavior observations within daily protocols, which is an improvement in utilizing the pig as an indicator of its status. Human observation is limited to sporadic viewing of many animals, and relevant trends are likely to be missed. An automated tool to evaluate behavior would alleviate some of the human workload for behavior observations and provide a more complete dataset from which to derive trends and ultimately valuable decisions. Existing tools and technologies have numerous limitations for practical deployment on commercial farms, including computing resources and processing time for real-time analysis. There is an opportunity to explore simplified computer vision techniques for observing pigs that yield sufficient information to provide valuable insights. Five computer vision techniques were implemented using three categories to evaluate the application in a simple behavior analysis. Each category represents a collection of behaviors representing a similar animal status for a group of pigs together in a pen. The automated determination of category was accomplished using machine learning and mathematical techniques with frames extracted from a set of labeled images from a commercial finishing pig farm. The image labels divide the image set into three activity categories based on behavior (categories 1-3), based on written definitions, assigned by a trained reviewer. Frames were manually separated and labeled, according to the input format for each model technique. Computer vision models with YOLOv8 and TensorFlow had activity level predictions of 86% and 79% overall accuracy, respectively, with the models having more mislabels for Category 2 and 3. The mathematical technique using Mean Square Error (MSE) correctly distinguished between activity categories 1 and 3 (P=0.000297), and both Root Mean Square Error (RMSE) and Structural Similarity Index Measure (SSIM) did not result in distinctly different categories.

Graduation Semester

2025-05

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/129666

Copyright and License Information

Chapter 4 contents are proprietary and are being submitted to the office of technology management

Owning Collections