Full text
| |
| Author(s): |
Paula de Freitas Curti
Total Authors: 1
|
| Document type: | Master's Dissertation |
| Press: | Pirassununga. |
| Institution: | Universidade de São Paulo (USP). Faculdade de Medicina Veterinária e Zootecnia (FMVZ/SBD) |
| Defense date: | 2025-05-19 |
| Examining board members: |
Ricardo Vieira Ventura;
Cristian Marlon de Magalhães Rodrigues Martins;
Guilherme Gomes da Silva
|
| Advisor: | Ricardo Vieira Ventura; Francisco Palma Rennó |
| Abstract | |
This document presents the activities and results of a master′s research project aimed at applying Computer Vision (CV) techniques to identify locomotor alterations and estimate respiratory rate (RR) in dairy cows. For locomotion analysis, 162 lateral-view videos of 67 Holstein cows were recorded immediately after milking. From a sample of 40 videos from 40 different animals, an algorithm was developed to automatically detect 14 body keypoints using the YOLOv8 architecture. The model achieved an accuracy of 99% and an average Object Keypoint Similarity (OKS) of 99% when all keypoints were simultaneously visible. This algorithm was then applied to the full dataset to extract locomotion-related features based on keypoint coordinates. Locomotion scores were assigned by two independent evaluators and used as a reference for evaluating the predictive model. In the final stage, despite the limited sample size, a classification algorithm was tested to detect locomotor problems without aiming to estimate herd-level prevalence, which may be addressed in future research with a larger sample. Due to class imbalance in the locomotion scores, the Random Forest classifier was used in conjunction with the SMOTE technique to balance the classes. A combinatorial attribute analysis was conducted to identify the best variable combinations for classification. The best-performing feature set included back area and the stride lengths of the left forelimb and hindlimb, yielding precision values of 0.1818, 0.5385, 0.9231, and 0.1667 for scores 1, 2, 3, and 4, respectively. These results highlight the impact of the small dataset and class imbalance, indicating the need for more robust datasets to validate the feature combinations most indicative of locomotor disorders. For RR estimation, 688 dorsal-view videos of 84 different Holstein cows were manually annotated and analyzed. Three segmentation approaches were tested in the CV model, differing in the region of interest within the image: Center Crop (CC), Full Bounding Box (FBB), and Corner Bounding Box (CBB). The CBB approach yielded the lowest errors, with a mean absolute percentage error (MAPE) of 14.04% and root mean square error (RMSE) of 4.93. The CC and FBB approaches resulted in MAPE values of 14.75% and 15.81%, and RMSE values of 5.41 and 5.84, respectively. These findings demonstrate that CV applied to all three image regions is technically viable for estimating RR in dairy cows. (AU) | |
| FAPESP's process: | 23/02851-0 - Computes Vision Technologies Applied to Lameness Detection in Dairy Cattle |
| Grantee: | Paula de Freitas Curti |
| Support Opportunities: | Scholarships in Brazil - Master |