ORP-Byte: A multi-object tracking method of pigs that combines Oriented RepPoints and improved Byte

The use of the multi -object tracking (MOT) algorithm in pig farming is profitable for identifying individual pigs automatically and monitoring their health status. However, tracking pigs in complex scenarios remains a challenge because of occlusion, overlapping, and shape deformation. To enhance the adaptability of tracking technology for group -housed pigs and to reduce pig identity switching (IDSW), we propose an MOT method based on a rotated bounding box detector. We named this method Oriented RepPoints (ORP)-Byte, which combines the Oriented RepPoints detector and the improved Byte algorithm. Based on the Byte algorithm, we designed an additional matching process, namely, the centre distance matching mechanism, which reduced the IDSW caused by the failure of the detector to locate pigs in time due to high-speed movement or deformation of pigs. To further our work, we constructed two detection datasets, two tracking datasets that contain daytime and nighttime images and annotations, and a public video dataset to launch a series of experiments based on them. The experimental results showed that ORP-Byte achieved an MOT accuracy (MOTA) of 99.8%, an IDSW of 16, an identity F1 -score (IDF1) of 91.6% and an average precision of ORP of only 82.4%. Compared with the SORT algorithm based on the Faster RCNN model, our proposed method displayed substantial improvements in MOTA and IDF1 by 1.2% and 28.5%, respectively, with a significant decrease of 86% in IDSW. In addition, even on the publicly available videos with a low frame rate, MOTA and IDF1 reached 96.0% and 94.1%, respectively and IDSW was only 5. ORP-Byte is effective and robust in the complex breeding scene, and it provides technical support for long-term, automatic, and non -contact pig monitoring.