IAE-YOLOv9: Advancing Automatic Wildlife Detection and Its Practical Application in Pantholops hodgsonii Monitoring for Biodiversity Conservation

Pantholops hodgsonii is a keystone species in plateau ecosystems, but its high-altitude, harsh habitat imposes significant challenges for traditional field surveys. This study presents an enhanced IAE-YOLOv9 model that integrates the Involution operator, AKConv lightweight architecture, and EMA multi-scale attention mechanism to improve detection performance in complex environments. Specifically, Involution enlarges the receptive field to strengthen small-object detection, AKConv dynamically adjusts kernel parameters to enhance sensitivity to small targets, and EMA refines boundary localization and classification through multi-scale feature fusion. Combined with infrared-triggered cameras, the system enables real-time monitoring, accurate detection, and automatic counting of Pantholops hodgsoni. On the self-constructed Sichuan-Gansu Wildlife Infrared Dataset (SGWID), IAE-YOLOv9 achieved 94.8% precision (95% CI: 94.2-95.1%), 95.2% mAP@0.5 (95% CI: 94.4-95.5%), and 91.2% recall (95% CI: 90.8-91.5%), outperforming existing models. When deployed along migration corridors, the system obtained 89.81% detection accuracy (95% CI: 88.34-91.12%) and 88.28% counting accuracy (95% CI: 87.45-89.12%) on the Hoh Xil Wildlife Camera Trap Dataset (HXWCTD), closely matching manual counts. This approach effectively supports population monitoring, migration assessment, and conservation decision-making, demonstrating the promise of integrating infrared cameras with deep learning in high-altitude regions.