Feeding behavior recognition of group-housed pigs based on pose estimation and keypoint features discrimination

• A feeding behavior recognition method based on pose estimation and keypoint features discrimination was proposed. • A Pig-HRNet pose estimation network was designed to effectively solve the problem of misdetection of occluded keypoints. • The proposed method can be used to distinguish between Feeding Behavior (FB) and Non-Nutritive Visiting Behavior (NNVB). • A group-housed pigs feeding behavior dataset was released for public research. In the field of intelligent sensing for smart animal husbandry, accurate recognition of feeding behavior in group-housed pigs is crucial for achieving precision farming and improving pig welfare. Currently, pig feeding behavior recognition relies on detection boxes-based methods, which are difficult to exclude Non-Nutritive Visiting Behavior within the feeding zone. To precisely recognize the feeding behavior of group-housed pigs, this study proposes a feeding behavior recognition method based on pose estimation and keypoint features discrimination. Firstly, Pig-HRNet is designed to estimate the pose of group-housed pigs, in which a Context Transformer (COT) attention module is specially introduced to detect the keypoints of pigs more accurately under crowded conditions. Secondly, by analyzing the correlation between keypoints and feeding zone, group-housed pigs are divided into visiting the feeding zone and Non-Feeding Behavior (NFB). For visiting the feeding zone, the behaviors are further categorized into Feeding Behavior (FB) and Non-Nutritive Visiting Behavior (NNVB). The experimental data of group-housed pigs were collected in commercial pig farms, including a total of 1400 video frames. Experimental results show that the Pig-HRNet model achieves an average precision (AP) of 97.1% in estimating pig poses. Compared to other pose estimation network models such as KAPAO, HigherHRNet, DeepLabCut, and HRNet, the detection AP improved by 69.0%, 16.3%, 12.3%, and 0.5%, respectively. The feeding behavior recognition method proposed in this paper achieved precision and recall rates of 98.8% and 99.9%, respectively. The relevant results indicate that the proposed feeding behavior recognition method performs well, while also meeting the requirement for accurately estimating pig poses under crowded conditions. The feeding behavior dataset established in this paper has been shared on https://github.com/IPCLab-NEAU/Group-housed-pigs-Feeding-Behavior-Recognition for use by the precision animal husbandry research community.