UD-YOLOv5s: Recognition of Cattle Regurgitation Behavior Based on Upper and Lower Jaw Skeleton Feature Extraction

Rumination is a crucial indicator for evaluating the health status of ruminants. However, Traditional contact devices, including ear tags and pressure sensors, may raise animal welfare concerns while detecting rumination behavior. Deep learning provides robust support for non-contact rumination behavior recognition through the training of datasets using neural networks. This paper presents a novel bovine rumination recognition method called UD-YOLOv5s, which combines YOLOv5s with upper and lower jaw skeleton feature extraction techniques. Firstly, a skeleton feature extraction method is proposed for the upper and lower jaws based on skeleton heatmap descriptors and the Kalman filter algorithm. Secondly, the UD-YOLOv5s method is designed for rumination recognition. Finally, a self-built bovine rumination dataset is used to compare the performance of three deep learning techniques, including MEAN-SHIFT, MASK-RCNN, and YOLOv3, with correlations. The results of the ablation experiment demonstrate that UD-YOLOv5s achieves a precision of 98.25%, recall of 97.75%, and a mean average precision of 93.43%. To ensure fairness, we conducted generalization performance-e evaluation in a controlled experimental environment, which showed that UD-YOLOv5s converge faster than other models while maintaining comparable recognition ac-curacy. Furthermore, our study demonstrates that when convergence speed is equal, UD-YOLOv5s outperforms other models in terms of recognition accuracy. Our findings provide robust support for the identification of cattle rumination behavior.