Optimized Real-Time Object Detection on Raspberry Pi 5 Using YOLOv5n/Openvino for VTOL

Abstract

This study evaluates a real-time object detection system on Raspberry Pi 5 for vision-based autonomous navigation of VTOL drone platforms. A lightweight YOLOv5n model was trained at 640×640 pixels and optimized by converting it to OpenVINO Intermediate Representation (IR) format with a reduced input size of 416×416 pixels to improve CPU inference speed on edge hardware. The dataset consists of three object classes (baskets, cubes, and tori) comprising 4,575 images with 10,260 annotated instances (approximately 3,420 annotations per class), collected in various environmental conditions using the Roboflow platform. Experiments were conducted on a VTOL drone platform under controlled and tethered conditions using a USB camera connected to the Raspberry Pi 5. Results show the system achieves [email protected] of around 0.992 and a maximum F1-score of around 0.992 at a threshold of 0.43, while the confusion matrix indicates high classification accuracy across all classes. The system achieves an average throughput of approximately 18 FPS and inference time of approximately 55 ms per frame across 2,160 outdoor test frames, with moderate CPU and memory usage. This demonstrates that YOLOv5n with OpenVINO on the Raspberry Pi 5 enables high-performance real-time multiclass detection, providing a strong foundation for drone navigation systems development.