Vision-guided tracking and detection using the YOLOv5 model on a logistic delivery fixed-wing UAV

Abstrak

Deep learning technologies utilizing Convolutional Neural Networks (CNNs) have advanced the development of autonomous systems, particularly in the exploration of hazardous environments. This study integrates the YOLOv5 object detection model with a fixed-wing Unmanned Aerial Vehicle (UAV) to identify simulated 5x5-meter orange marker dropping kits deployed in inaccessible disaster zones. Experiments were conducted at altitudes of 45 m, 75 m, and 100 m above sea level to assess real-time detection accuracy and terrain-mapping efficiency. The system achieved a mean Average Precision (mAP) of 88% across varying altitudes, demonstrating robust performance despite environmental challenges such as false positives from similarly colored rooftops. Computational efficiency tests were performed on the Jetson Nano platform using the TensorRT engine to accelerate object detection model inference on NVIDIA GPUs. Lighting variability significantly impacted detection reliability, resulting in a reduced mAP under suboptimal illumination. To enhance precision, post-processing filters and parameter optimizations were applied, improving the balance between detection sensitivity and specificity. These findings underscore the potential of YOLOv5-enabled UAVs for rapid, high-accuracy aid localization in disaster scenarios, although adaptive threshold tuning remains critical to address environmental variability in operational settings.