DeMVpp-YOLO: A lightweight pig behaviour detection model for improving pig health management in farrowing pens

The behaviours exhibited by pigs are closely tied to their health and welfare status, making it crucial to monitor them for effective management. Automatic real-time detection methods can significantly improve pig health management and reduce the workload of caretakers, particularly in farrowing pens where piglets and sows are vulnerable. In recent years, computer vision, particularly empowered by deep learning, has gained traction for animal behaviour recognition. However, existing deep learning methods either require high computational resources or lack the detection accuracy needed for real-world applications. To balance the detection precision and model complexity, we propose a novel lightweight network called DeMVpp-YOLO. We improve the backbone and neck of the Transformer-based MobileViTv2 and PPyolov5 models, respectively, to reduce the model size and enhance deep feature extraction. Additionally, we employ a decoupled head to separate the classification and regression tasks, thereby accelerating the training process. In our pig detection experiment, the detection mean Average Precision (mAP) for piglet and sow behaviours reached 93.1% in the ablation study, demonstrating significant improvement in detection compared with YOLOv-5. Our experiments successfully identified four postures including standing, left lying, right lying, and sternal lying for piglets and sows, with an average accuracy rate of 96.3%. Furthermore, to assess the activity levels of piglets and the sow during the early stages of development in the farrowing pens, we utilized the model to characterize the pig posture patterns on two different days. © 2024 11th European Conference on Precision Livestock Farming. All rights reserved.