2026-04-13T18:26:45Zhttps://recercat.cat/oai/request

oai:recercat.cat:2117/4139712025-07-17T15:43:58Zcom_2072_1033col_2072_452951

Animal tracking and identification in smart farms Compte Prades, Albert Àrees temàtiques de la UPC::Informàtica::Automàtica i control Animal welfare Animal handling Computer vision Identificació d'animals Seguiment d'animals Granges Intel·ligents Re-identificació LehmanPigBoy-33 SORT Visió per computadora Seguiment Multi-Objecte Faromatics SAM YOLOv8 Base de dades Animal Identification Animal Tracking Smart Farm Re-Identification LehmanPigBoy-33 SORT Computer Vision Multi-Object Tracking Faromatics SAM YOLOv8 Dataset Animals--Protecció Animals--Manipulació Visió per ordinador With the rising demand for affordable animal products and the increasing concerns regarding animal welfare, it is becoming imperative the development of more efficient livestock farming techniques that not only enhances productivity but also ensures that ethical standards are met. That is specially relevant on the sector of pig farming, as pork production is among the most demanding markets in the agriculture sector, contributing on the worldwide economy. This prominence underscores the necessity for specialized monitoring systems tailored to their unique behaviors and physiological needs. The purpose of this thesis is the design, development and test of the identification and tracking components of a tracking-by-detection Multi-Object Tracking (MOT) algorithm as part of a collaboration between the HuPBA research group and Faromatics (AGCO). More specifically, this project is split into two folds: the development and creation of a pig gallery dataset, called LehmanPigBoy-33 Dataset, and the design and experimental test of the re-identification and tracking components. During the creation of the dataset, several videos of a pig farm in the United States were recorded using a mobile robot, and different annotation pipelines were created to improve the annotation of the identities. Our proposed model comprises four components: a detector module based on YOLOv8, a re-identification module based on ResNet-50, a tracker module based on the SORT tracker, and a voting module that implements a voting scheme between the IDs produced by the re-identification module and the trajectories of the tracker module. This work focused exclusively on the latter three components. The re-identification module achieved a test accuracy of 85% on a dataset of 33 identities, while the voting scheme, based on maximum confidence voting, improved the test accuracy of a recorded video from 90% to 98%. 2024-05-15 Master thesis Open Access Universitat Politècnica de Catalunya