Ahmad Yogi Kulumul Ilham – Poster

Monitoring humans and wildlife using UAV mounted thermal cameras is crucial for anti poaching and conservation in low light, wide area settings. This study benchmarks four modern YOLO detectors YOLOv8, YOLOv9, YOLOv10, and YOLOv11 fine tuned on thermal UAV imagery from the BIRDSAI / LILA Conservation Drones dataset covering five classes (human, elephant, lion, giraffe, unknown). We standardize the training pipeline at 640 pixel resolution and evaluate models with common detection criteria and runtime analysis on an NVIDIA A100. Qualitative and quantitative comparisons indicate that YOLOv10 delivers the highest throughput for real-time deployment, while YOLOv11 offers the most balanced trade-off between detection quality and speed across diverse scenes. Error analysis highlights persistent challenges with small, low contrast targets and background confusion typical of thermal imagery. The results provide a practical guide for selecting detectors under real-time constraints, and we include curated precision recall curves, confusion matrices, and exemplar detections to support future development of thermal UAV monitoring for wildlife conservation and anti poaching.