Transformer or CNN? Benchmarking Real-Time Detection Transformer and YOLOv8 for Small UAS Autonomous Landing

Small Unmanned Aircraft Systems (UAS) are increasingly used in commercial package delivery, requiring robust object detection capabilities for safe and reliable landing in dynamic, off-nominal environments. Deep learning based computer vision models have shown promise in achieving high-accuracy object detection and show promise as key components to the UAS autonomy stack. In recent years, transformer architectures have emerged as powerful alternatives to convolutional models, offering strong representation capabilities for object detection tasks. This paper compares two state-of-the-art object detection models for aerial-view object detection: the convolutional neural network (CNN) YOLOv8 and Real-Time Detection Transformer (RT-DETR) models. Both models are trained and tested on the VisDrone aerial imagery dataset, which provides relevant data for detecting variably scaled and dense urban objects commonly encountered during UAS descent and landing. Performance is assessed using standard computer vision detection accuracy metrics and inference latency to assess real-time feasibility. To further evaluate the reliability of each model’s predictions, we apply the conformal prediction framework, enabling statistically valid estimates of classification uncertainty. Our results provide insight into the trade-offs between detection performance, inference speed, and uncertainty calibration, offering guidance for selecting appropriate object detection models in safety-critical UAS landing applications.