Exploring Deep Learning for Detection of Poultry Activities — Towards an Autonomous Health and Welfare Monitoring in Poultry Farms

Abstract

The health condition of poultry significantly affects egg production, meat quality, and reproduction. Behavioral activities such as feeding patterns can be indicators of their current welfare. However, assessing through on-site observation is tedious, time-consuming, possibly biased, and can induce stress to the birds. Hence, employment of an autonomous surveillance system that can continuously and noninvasively monitor the poultry behaviors is the most viable approach. In this study, detection of quail activities: eating, drinking, and roaming, is administered using computer vision (CV) and deep learning (DL). Four DL models, YOLOv5, YOLOX, Faster R-CNN, and EfficientDet, were explored to detect quail activities in cages. The three models YOLOv5, YOLOX, and Faster R-CNN, achieved an average precision (AP) of 85.52, 79.31, and 74.28, respectively. For the EfficientDet model, the training was evaluated using total loss. A total loss of 0.1616 was achieved at 10,000 iterations. All the DL models performed impressively in detecting quail activities in cages. This study contributes to the development of an intelligent health assessment system for poultry.