DroNet: Efficient convolutional neural network detector for real-time UAV applications

Kyrkou, Christos; Plastiras, George; Theocharides, Theocharis; Venieris, Stylianos I.; Bouganis, Christos-Savvas · 2018 · OpenAlex-citations

DOI: 10.23919/date.2018.8342149

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of deploying real-time object detection on Unmanned Aerial Vehicles (UAVs), which are constrained by limited power, memory, and processing capabilities. While deep learning offers high accuracy for tasks like vehicle detection in traffic monitoring and emergency response, conventional architectures such as Faster R-CNN are computationally heavy, and GPUs introduce prohibitive power consumption and latency due to batch processing requirements. Cloud-based solutions are unsuitable for latency-sensitive or remote operations. Consequently, the authors propose DroNet, a lightweight, single-shot convolutional neural network (CNN) designed specifically for efficient on-board vehicle detection from aerial imagery. The study employs a holistic approach involving data collection, model design, and hardware optimization. The authors constructed a dataset of 350 images containing approximately 5,000 vehicles, sourced from satellite imagery, web repositories, and UAV footage to ensure diversity in illumination, viewpoint, and occlusion. Using the Darknet framework, they trained and evaluated four CNN architectures—SmallYoloV3, TinyYoloVoc, TinyYoloNet, and DroNet—derived from the Tiny-YOLO model. The design process involved pruning layers, reducing filter counts, and varying input image sizes to balance computational load with detection accuracy. Performance was assessed using Intersection over Union (IoU), sensitivity, precision, and frames-per-second (FPS), combined into a weighted score that prioritized speed. Experimental results demonstrate that DroNet achieves a significant performance improvement over baseline models. On an Intel i5 CPU, DroNet was 30 times faster than TinyYoloVoc with only a minimal drop in IoU. When deployed on embedded platforms suitable for UAVs, DroNet maintained an overall accuracy of approximately 95%. Specifically, on an Odroid XU4 board, the model achieved 8–10 FPS, while on a Raspberry Pi 3, it operated at 5–6 FPS. The optimal input resolution was determined to be 512x512 pixels, maximizing the trade-off between speed and accuracy. The authors note that the 40x speedup over previous works makes the model viable for real-time applications, despite a slight increase in false detections due to reduced network complexity. The significance of this work lies in providing a practical, efficient architecture for resource-constrained UAV applications. By demonstrating that high-accuracy vehicle detection can be performed on low-power embedded processors without cloud dependency, DroNet enables real-time situational awareness for traffic management and search-and-rescue missions. The paper concludes that further optimizations, such as bitwidth reduction and expanded training datasets for additional object classes, could further enhance the utility of such systems in diverse operational environments.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• distraction detection algorithms