FTO-SORT: a fast track-id optimizer for enhanced multi-object tracking with SORT in unseen pig farm environments

초록

As the importance of animal welfare and agricultural automation continues to grow, advancements in real-time multi-object tracking technology within pig farm environments have become essential. In commercial farms (unseen datasets) differing from the trained dataset, various factors, such as light flares, object overlap, and ambiguous boundaries between the foreground and background, pose challenges that reduce detection accuracy. Although methods using large models or ensemble detectors exist, they often suffer from slower execution speeds. While other studies focus on improving accuracy on seen datasets, this can result in lower accuracy when applied to different farm environments. To address this, our study introduces a new data augmentation method called FBDA (Foreground-Background Separation and Data Augmentation) using SAM and LaMa to enhance performance without affecting execution speed. We also implemented a new weighting technique in the existing loss function to handle overlaps or ambiguous boundaries, termed OFBL (Overlap and Foreground-Background Difference Loss), which improves detection performance while maintaining YOLO's speed. To further improve tracking performance, we introduced a new module, FTO (Farm Track-id Optimizer), into the BoT-SORT (Bag of Tricks for MOT Simple Online and Realtime Tracking) model, resulting in FTO-SORT (Farm Track-id Optimizer with SORT). Using our proposed method, we achieved significant improvements in tracking performance on unseen datasets, increasing IDF1 from 75.1% to 90.2% with YOLOv8 and from 68.7% to 86.7% with YOLOv11, representing gains of 15.1% and 18.0%, respectively. Additionally, by removing the Re-ID module from FTO-SORT, the FPS on the TX2 board increased, achieving speeds that were approximately 10.3 times faster (from 0.6 to 6.2 FPS) and 11.1 times faster (from 0.6 to 6.7 FPS), significantly enhancing processing speed. We shared our tracking dataset at https://github.com/YuSeungHyun97/fto-sort for precision livestock farming research community.

키워드