ShuffleDet: Real-Time Vehicle Detection Network in On-Board Embedded UAV Imagery

Azimi, Seyed Majid · 2019 · Crossref

DOI: 10.1007/978-3-030-11012-3_7

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 introduces ShuffleDet, a computationally efficient deep learning network designed for real-time vehicle detection in unmanned aerial vehicle (UAV) imagery. The research addresses the challenge of deploying object detection algorithms on embedded mobile platforms with limited processing power and strict energy constraints. While existing convolutional neural networks (CNNs) like Faster R-CNN and YOLO offer high accuracy, they are too computationally demanding for on-board UAV applications. Conversely, lightweight networks often sacrifice accuracy. The authors aim to bridge this gap by developing a method that maintains competitive detection performance while ensuring low computational complexity and high inference speed. The proposed ShuffleDet architecture combines ShuffleNet as the backbone feature extractor with a modified Single Shot MultiBox Detector (SSD) framework. To reduce computational load, the network utilizes channel shuffling and grouped convolutions. To address the specific challenges of aerial imagery, where vehicles appear small and exhibit varied geometric shapes, the authors incorporate modified Inception modules and Domain Adapter Blocks (DABs). The Inception modules use depthwise convolutions to capture features at multiple scales efficiently. The DABs integrate deformable convolutions to adapt pre-trained weights from terrestrial datasets to the aerial domain and to better model vehicle geometries. The model was trained on the ImageNet-2012 dataset and evaluated on the CARPK and PUCPR+ datasets, with images resized to 512x512 pixels. Implementation was conducted using Caffe, with training on an Nvidia Titan XP GPU and evaluation on an NVIDIA Jetson TX2 embedded device. Experimental results demonstrate that ShuffleDet achieves a computational complexity of only 3.8 GFLOPs, significantly lower than state-of-the-art methods like Faster R-CNN (118.61 GFLOPs) and YOLO (26.49 GFLOPs). On the CARPK dataset, ShuffleDet achieved an RMSE of 38.46, outperforming MobileNet-SSD (65.24 RMSE) and approaching the performance of heavier models. On the PUCPR+ dataset, it achieved an RMSE of 49.68. Crucially, the network runs at 14 frames per second (FPS) on the Jetson TX2, compared to 1–8 FPS for competing architectures. Ablation studies confirmed that the modified Inception modules and DABs significantly reduced RMSE, validating their role in enhancing small object detection and domain adaptation. Generalization tests on the 3K-DLR-Munich dataset further showed consistent performance with an inference time of 524 ms per high-resolution image. The significance of this work lies in its demonstration that high-accuracy, real-time vehicle detection is feasible on resource-constrained embedded hardware. By achieving 14 FPS on an edge device, ShuffleDet enables immediate applications such as traffic monitoring, search and rescue, and urban management without reliance on cloud-based processing. The study highlights that strategic architectural modifications, such as deformable convolutions and channel shuffling, can effectively balance the trade-off between speed and accuracy, making it a viable solution for autonomous UAV systems.

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