Real-Time Livestock Activity Monitoring via Fine-Tuned Faster R-CNN for Multiclass Cattle Behaviour Detection

Automated cattle activity detection plays a pivotal role in modern livestock management, significantly impacting animal welfare and operational efficiency. This paper introduces an automated approach for cattle activity detection using advanced deep learning-based architecture named Faster-RCNN. The deep learning model addresses the simultaneous detection and precise localization of three primary cattle activities in the input image: standing, lying, and walking. The methodology involves fine-tuning a pre-trained model using a dataset collected from a real-time barn environment at Hartpury University Farm. Overall, the proposed approach is based on data pre-processing and fine-tuning steps. Data augmentation techniques, such as random cropping, flipping, and rotation, ensure dataset diversity. This enriches the model’s generalization ability across various lighting conditions and cattle orientations. The fine-tuning process adapts a pre-trained model, initially trained on a general object detection dataset. We adjust the model’s architecture to the subtleties inherent in different cattle activities through training on our custom cattle activity dataset. This process ensures the model’s significance in accurately detecting and classifying the distinct behaviours of cattle. Experimental results demonstrate the model’s effectiveness in identifying and localizing cattle activities. The model correctly predicted of standing, lying, and walking events with accuracy rate of 0.94, 0.92 and 0.89 respectively.