GAM-YOLOv7-tiny and Soft-NMS-AlexNet: Improved lightweight sheep body object detection and pose estimation network

Abstract

Intelligent livestock farming has been a major focus of attention in recent years. Using deep learning to assist livestock management can effectively reduce labor loss and improve management efficiency. Lightweighting plays a key role in the deployment and practical use of deep learning models, and most existing sheep-based deep learning models do not focus on this, which has become a major factor limiting the development of smart sheep farming. Therefore, in this paper, first, a GAM-YOLOv7-tiny neural network model for object detection of sheep was investigated. The size of the model reached 3.5G, which was a reduction to 26.3% of the original size, the FLOPS was reduced by 74.1%, the experimental result reached 96.4% of mAP and the FPS reached 88.232 on an RTX 1650Ti. Second, a Soft-NMS-AlexNet neural network model for key point recognition of sheep bodies was investigated with a model size of 1.97G, and the final experimental results achieved 82% AP and 190.86 ± 23.97 FPS. Finally, we completed the behavior recognition of the standing and lying posture of sheep using the pose estimation model, which provides a research solution for performing behavioral monitoring and giving early warnings for diseases for sheep.