Deep Transfer Learning for Classification of Rabbit Behaviour Using Publicly Available Datasets

Deep learning has emerged successfully as a utilised technology for animal monitoring and classification. However, there exists limited research when it comes to rabbit monitoring and classification. This study presents real time models for detection and tracking of rabbits, along with a classification model for their behaviour. The detection model was trained on a dataset obtained from Robflow and YouTube, while the classification models utilised over 1100 curated images representing seven different physical behaviours of rabbits: digging, humping and mating, eating, periscoping, jumping and binkying, laying, and flopping and loafing. For object tracking, transfer learning was applied using a pretrained ResNet18 model for feature extraction, and YOLOv8 was employed for tracking, supported by ByteTRACK for consecutive frame tracking. Additionally, VisionTransformer, ResNet18, and EfficientNet models were finetuned to adapt their representations for behaviour classification. The experimental result demonstrated that the finetuned VisionTransformer model achieved an accuracy of 86%, outperforming ResNet18 and EfficientNet models with accuracies of 76% and 80% respectively. In the object detection experiment, the YOLOv8 model achieved a mean Average Precision (mAP50) score of 0.96. These results indicate the effectiveness of the proposed approach for accurate detection, tracking, and classification of rabbit behaviour in real-time.