PKL-Track: A Keypoint-Optimized approach for piglet tracking and activity measurement

Abstract

This study proposes an efficient and accurate multi-object tracking method for piglets (Piglet Keypoints and L2 Distance Tracking, PKL-Track) to achieve piglet state monitoring and activity quantification. The proposed method employs the improved YOLOv11s-Pose model for target and keypoint detection, utilizing the relative positions of piglet bounding boxes to refine keypoint regression while optimizing the detection head to enhance model efficiency. To address challenges such as occlusion and target crowding, the BoT-SORT algorithm was improved by incorporating keypoint and bounding box information to refine matching distances, supplemented by normalized Euclidean distance to expand matching range. Experiments were conducted using video data from 31 piglet pens, constructing a dataset containing targets, keypoints, and tracking annotations for testing. Results demonstrated that the improved YOLOv11s-Pose model achieved an average precision of 98.5 % for object detection and 98.0 % for keypoint detection, with a detection time of 5.0 ms per frame. For multi-object tracking tasks, short frame intervals (5 frames) achieved 84.3 % HOTA, 99.1 % MOTA, and 91.5 % IDF1, significantly reducing ID switches. Activity quantification experiments based on tracking results revealed a relative error of only 2.36 % in group activity measurement, accurately reflecting piglet activity levels. The proposed method demonstrates excellent performance in multi-object tracking and activity quantification, providing key technological support for behavior monitoring and piglet health assessment in precision livestock farming.