PigTracker: Sequential-vertices-based tracking framework for intelligent pig farming

Multi-object tracking (MOT) is essential in intelligent pig farming, playing a significant role in automated identification, behavioral analysis, and health monitoring of pigs. However, existing MOT methods for pig farming face challenges of low-accuracy and weak-generalization under occlusion, crowded and overlapping scenarios. In this work, we propose a MOT framework, namely PigTracker, for overcoming these challenges. The proposed PigTracker consists of two collaborative modules: the detection module and the tracking module. In the detection module, we propose an efficient sequential-vertices oriented bounding box (SV-OBB) detector with modulated loss function. By incorporating an attention mechanism module, our method generates orientation-aware detection boxes for individual pigs. In the tracking module, we augment the Kalman filter’s state vector with orientation information from the detection stage. Building upon this, we design a temporal-window-based outlier detection and correction mechanism, effectively rectifying false orientation originating from the detection phase. Experimental results demonstrate that our PigTracker framework achieves state-of-the-art (SOTA) performance on multi-pig tracking task under crowed environment. Our sequential-vertices orientation-aware detector achieves a Average Precision (AP 50 : 95 ) of 89.0% and a Directional Mean Absolute Error (D-MAE) of 2.15° for orientation estimation. Tracking experiment attains a Higher Order Tracking Accuracy (HOTA) of 93.44%, surpassing the baseline HBB + ByteTrack (HOTA: 89.79%) by 3.65 points. In addition, the detection and tracking pig farming dataset build on this work will be publicly available for academic research, which can be accessed via https://github.com/linwjscau/PigTracker_MOT-Framework-For-Pig-Farming.git .