Reliable plateau livestock monitoring in complex environments via RGBT fusion tracking: a case study in Guide County, Qinghai Province

Precision livestock farming requires robust, non-contact monitoring in high-altitude pastures, where dense herds, target occlusion, illumination variation, and temporary degradation of either visible or thermal imagery complicate multi-object tracking. This study presents a UAV-based RGBT tracking framework for plateau livestock in Guide County, Qinghai Province. Because the field imagery contains animals with cattle-like and yak-like visual characteristics and no species-level labels were assigned, the monitoring target is defined as plateau livestock and the self-collected benchmark is referred to as the Guide-Livestock dataset. The framework combines a parameter-efficient Parallel Adaptive Residual Network (PARN) inserted into a frozen ResNet-50 backbone, dual-stream RGB and thermal encoding, staged detection and joint Transformer decoders, a KAN-GRU long-term memory with temporal prediction, StrongSORT-based association, and modal dropout for sensor degradation. On Guide-Livestock, the full model achieved HOTA = 71.5%, MOTA = 87.6%, IDF1 = 82.3%, AssA = 70.8%, and 58 FPS. On the public RGBT234 benchmark, it achieved HOTA = 65.3%, MOTA = 82.3%, IDF1 = 81.4%, AssA = 65.4%, and 53 FPS. Controlled ablations show that PARN alone improves HOTA, IDF1, and AssA by 3.2, 3.3, and 3.4 percentage points, respectively, while KAN-GRU improves the same metrics by 3.8, 5.0, and 4.0 points over the no-memory baseline. The results demonstrate the potential of RGBT fusion and long-term temporal modeling for real-time monitoring of free-ranging plateau livestock under occlusion, low illumination, and partial modality failure.