Objective To address the issue of pig trajectory loss caused by prolonged severe occlusion, we propose a multi-object tracking method for pigs based on oriented bounding box (OBB) annotation to enhance the tracking performance of pigs in occluded scenarios.

Method First, the YOLOv11n baseline model was improved by integrating a C3k2_DualConv convolutional network and a bidirectional feature pyramid network (BiFPN), thereby constructing a pig detection model named YOLO-DB. Second, building upon the BoT-SORT tracking algorithm, an enhanced matching strategy was developed using a trajectory frame analysis mechanism, incorporating posture consistency features before and after occlusion, and adopting a probabilistic intersection over union (ProbIoU)-based target matching mechanism optimized for OBB annotation. Finally, the YOLO-DB algorithm and improved BoT-SORT tracking algorithm were integrated to recover lost trajectories.

Result The YOLO-DB algorithm achieved precision, recall and mAP₅₀ of 96.5%, 95.6% and 97.3%, respectively, representing increases of 2.7, 1.2 and 1.4 percentage points compared to the baseline model, while its parameter count was reduced by 2.4%. The improved tracking algorithm attained scores of 82.4% for higher-order tracking accuracy (HOTA), 97.7% for multiple object tracking accuracy (MOTA), and 89.1% for identification F1 score (IDF1), outperforming the baseline model by 0.9, 1.3, and 5.4 percentage points, respectively. Additionally, identity switches (IDS), false positives (FP), and false negatives (FN) were significantly reduced.

Conclusion The proposed algorithm effectively resolves the problem of pig trajectory loss caused by prolonged occlusion and significantly enhances tracking performance in occluded scenarios. It provides an efficient and reliable technical solution for intelligent management in large-scale pig farms.