Pig Aggression Tracking and Analysis Application Based on A RPMeMOTR Method

The management of pig aggression in group-housed environments is crucial for ensuring animal welfare and optimizing production efficiency in the global agricultural industry. Due to rapid pig movement and frequent occlusions during aggression activities, the key challenges for pig tracking and aggression analysis using multi-object tracking (MOT) are low tracking accuracy and error identity-switching (IDs) problems. To address the above challenges, this study proposes an RPMeMOTR approach by combining the RKNet model and the P-ids algorithm based on MeMOTR for pig aggression tracking and analysis. Firstly, we create a new backbone structure named RKNet to enhance aggression feature extraction and representation learning ability. Then, we design the P-ids algorithm to ensure stable tracking by preserving and leveraging historical trajectory data. Finally, we develop an analysis algorithm to quantify the intensity of pig aggression to enable alerts for farm staff. In public database, RPMeMOTR achieved an average Higher Order Tracking Accuracy (HOTA) of 78.2%, Multi-Object Tracking Accuracy (MOTA) of 93.8%, identification F1 Score (IDF1) of 94.9%, and IDs of 7. Compared to MeMOTR, it demonstrated significant improvements in HOTA, MOTA, IDF1, and IDs metrics with 5.0%, 6.6%, 7.0% increase, and 11 reduces. And using the RPMeMOTR method, compared to OC-SORT, there was a 4.1% increase in HOTA, a 7.4% increase in IDF1 and 28 reduces in IDs, while compared to ByteTrack, its improvements were 4.8% in HOTA,8.6% in IDF1 and 33 reduces in IDs. It also shows superiority over other methods in private datasets. These experimental results demonstrated that the proposed approach offered an effective solution for the accurate tracking and aggression behavior analysis of pigs in intricate farm environments, meeting the requirements for long-time monitoring and intervention.