Benchmark on Real-Time Long-Range Aircraft Detection for Safe RPAS Operations

Abstract

The growing market in Remotely Piloted Aircraft Systems (RPAS) and the need for cost-effective “Detect and Avoid (DAA)” systems are critical issues up to date towards enabling safe beyond visual line of sight (BVLOS) operations. In hopes of promoting earlier threat detection on DAA systems, we benchmark several object detection algorithms on multiple graphical processing units for the concrete DAA use case. Two state-of-the-art “real-time object detection” and “object detection” model sets are trained using our CENTINELA dataset, and their performances are compared for a wide range of configurations. Results demonstrate that one-stage architecture YOLO variants outperform ViT on all tested hardware in terms of mean average precision and inference speed despite their architecture complexity gap. Additional resources are available to the reader at https://github.com/fada-catec/detection-for-safe-rpas-operation.