Enhanced DeepSORT and StrongSORT for Multicattle Tracking With Optimized Detection and Re-Identification

Recently, labor shortages in the farming industry have increased the demand for automation. Object tracking technology has emerged as a critical tool for monitoring livestock through automated systems. This study focuses on tracking individual cattle using object detection and tracking algorithms. Data were collected noninvasively using cameras, and a tracking-by-detection (TBD) approach was adopted. The proposed framework introduces multiple enhancements optimized for cattle tracking. These enhancements include a comparison of five different bounding box regression losses to improve detection accuracy, modifications to the Kalman filter state vector for more accurate bounding box predictions, and adjustments to the feature vector distance metric in the re-identification algorithm. YOLOv9-t was used as the detector, whereas DeepSORT and StrongSORT served as trackers. Compared with the baseline, which uses DeepSORT, the proposed method achieved significant improvements in higher-order tracking accuracy (HOTA) by 4.1%, multiple object tracking accuracy (MOTA) by 1.08%, and identification F1 score (IDF1) by 5.12%, reaching values of 78.64%, 90.29%, and 91.41%, respectively, while reducing the number of ID switches (IDSW).