Pigs' BCS Estimation using Computer Vision and Deep Learning Approaches

Abstract

This research presents an end-to-end automated system for Body Condition Score (BCS) assessment in pigs using multi-modal RGB-D computer vision and deep learning. Manual BCS evaluation is subjective, labor-intensive, and inconsistent across large-scale farming op erations. To address these limitations, we developed a hybrid ensemble deep learning pipeline combining ResNet-50 and DenseNet-121 architectures with integrated depth information from Intel RealSense D435 cameras. The system was trained and validated on a dataset of 268 pigs across six pens collected from Virginia State University farms, with video streams captured as .bag files and converted to PNG images for analysis. Experimental results demonstrate that the multi-modal RGB-D approach achieves a 13.68% accuracy improvement over traditional RGB-only methods when evaluated using the en semble model. The hybrid ensemble achieves 84.18% accuracy using multi-image temporal aggregation across five architectures: ResNet-50, DenseNet-121, EfficientNetV2-S , Vision Transformer, and the proposed hybrid ensemble . Overall, the system achieves 84.18% multi-image classification accuracy. The proposed automated pipeline demonstrates the feasibility of objective and scalable livestock health monitoring, with potential productivity gains through improved nutri tional management. Future work will focus on expanding the dataset through multi-farm validation and integrating behavioral monitoring systems to enable more comprehensive an imal welfare assessment.