Full text
| |
| Author(s): |
Gabriel Pagin de Carvalho Nunes Oliveira
Total Authors: 1
|
| Document type: | Master's Dissertation |
| Press: | Pirassununga. |
| Institution: | Universidade de São Paulo (USP). Faculdade de Zootecnica e Engenharia de Alimentos (FZE/BT) |
| Defense date: | 2024-03-14 |
| Examining board members: |
Rafael Vieira de Sousa;
Adroaldo José Zanella
|
| Advisor: | Rafael Vieira de Sousa |
| Abstract | |
Recent studies in animal production demonstrate that body mass prediction serves as assistance to animal management and in the quality of produced animal protein. In this context, the construction and evaluation of a computer vision system for pig body mass prediction through point clouds and radio frequency identification (RFID) were pursued. For this purpose, RGB-D videos were captured using a depth camera installed in a pen containing 25 pigs (Landrace x Large White), 15 males and 10 females ,aged around 12 weeks, with body mass ranging from 29.5 kg to 93.5 kg, associated with RFID identification for each animal. After selecting the best videos, 721 point clouds were extracted. The height was extracted based on the average of the top 3% highest points on the back of each animal. A cut-off line was established to eliminate outlier points at the base of the point cloud by removing a percentage of the height in relation to the base. Using a Python script, the Convex Hull (CH) and Alpha Shape (AS) algorithms were employed to extract dimensional characteristics of the pigs\' dorsal region such as perimeter, surface area (3D area), projection of surface area (2D area), and volume, generating two datasets (CH dataset and AS dataset). Subsequently, computational models based on machine learning algorithms (decision tree (RT), random forest (RF), k-nearest neighbors (KNN), support vector regression (SVR), linear regression (LR), and artificial neural network (MLP) were trained and compared using a cross-validation process followed by statistical analysis. Attribute selection was performed through a Wrapper function and further statistical analysis. The statistical analysis revealed significant differences between the tested algorithms, except between KNN and SVR, and between SVR and MLP for the CH dataset, and between SVR and MLP for the AS dataset. The RF model achieved the best performance in both datasets. The AS method demonstrated better performance compared to the CH dataset. The RF model with the AS dataset was selected as the best model in this study, resulting in an R² of 97.77%. The Wrapper resulted in the exclusion of surface area (3D area) and volume without altering the model outcome in both datasets (CH and AS). The results suggest significant potential in creating more robust computational models for body mass prediction, opening doors for future developments with direct applicability in commercial pen environments. (AU) | |
| FAPESP's process: | 22/11652-8 - Computer vision applied to the monitoring of individual body mass of finishing pigs |
| Grantee: | Gabriel Pagin de Carvalho Nunes Oliveira |
| Support Opportunities: | Scholarships in Brazil - Master |