Accès libre

Lightweight Low-Altitude UAV Object Detection Based on Improved YOLOv5s

Haokai Zeng
Haokai Zeng
, 
Jing Li
Jing Li
 et 
Liping Qu
Liping Qu
   | 28 mars 2024
International Journal of Advanced Network, Monitoring and Controls's Cover Image
À propos de cet article
Article précédent
Article suivant
Citez
Publié en ligne: 28 mars 2024
Pages: 87 - 99
DOI: https://doi.org/10.2478/ijanmc-2024-0009
Mots clés
© 2024 Haokai Zeng et al., published by Sciendo
This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Figure 1.

ATD-YOLO Network Structure
ATD-YOLO Network Structure

Figure 2.

Framework of image measuring system [14]
Framework of image measuring system [14]

Figure 3.

C3F structural schematic diagram
C3F structural schematic diagram

Figure 4.

CARAFE upsampling calculation flowchart
CARAFE upsampling calculation flowchart

Figure 5.

EMA Attention Mechanism
EMA Attention Mechanism

Figure 6.

GSConv Module
GSConv Module

Figure 7.

VoVGCSP Module
VoVGCSP Module

Figure 8.

The positions of GSConv and VOVGCSP modules
The positions of GSConv and VOVGCSP modules

Figure 9.

Length and Width Distribution Chart of the Anti-Mini Drone Dataset
Length and Width Distribution Chart of the Anti-Mini Drone Dataset

Figure 10.

Samples of simple background from Anti-Mini Drone
Samples of simple background from Anti-Mini Drone

Figure 11.

Samples of complex background from Anti-Mini Drone
Samples of complex background from Anti-Mini Drone

Figure 12.

PR curves for various feature extraction modules (IOU=0.5)
PR curves for various feature extraction modules (IOU=0.5)

Figure 13.

PR curves for various feature extraction modules (IOU=0.5)
PR curves for various feature extraction modules (IOU=0.5)

Figure 14.

Model PR curve (IOU=0.5)
Model PR curve (IOU=0.5)

Figure 15.

PR curves of mainstream algorithms on the test set (IOU=0.5)
PR curves of mainstream algorithms on the test set (IOU=0.5)

Figure 16.

Object detection outcomes in diverse scenarios
Object detection outcomes in diverse scenarios

Figure 17.

Object detection outcomes in a consistent scenario
Object detection outcomes in a consistent scenario

Mainstream Algorithm Comparative Experiment Results

Module Params/106 GFLOP/G AP.5/% FPS
YOLOv3 Tiny 8.66 12.9 79.1 166.67
YOLOv5s 7.01 15.9 92.2 68.79
YOLOv7 Tiny 6.01 13.2 88.4 63.30
YOLOv8s 11.12 28.4 89.0 109.89
ATD-YOLO 5.23 11.0 92.8 75.35

Results of ablation experiments

YOLOv5s C3F EMA CARFE Slim-Neck Params/10 6 GFLOP/G mAP.5/% FPS
√ 7.01 15.8 92.2 68.79
√ √ 6.33 13.8 92.3 75.05
√ √ √ 6.38 14.1 92.7 69.83
√ √ √ √ 6.40 14.1 93.1 67.85
√ √ √ √ √ 5.23 11.0 92.8 75.35

Contrast experiment of attention module

Module mAP.5/% GFLOP /G Params/106 FPS
SE[24] 91.8 13.8 6.37 74.93
ECA[25] 92.2 13.8 6.34 74.37
CBAM[26] 92.3 13.8 6.37 72.63
CA[27] 91.2 13.8 6.36 72.87
EMA 92.7 14.1 6.38 69.83

Origin of the Dataset and Quantity of Images

Dataset Number of Images
Det-Fly 3893
Drone-vs-Bird 3959
Real World 1525
Multi-view drone tracking 3447
DUT anti-UAV 3639
Anti-UAV 2767

Experimental Setup Configuration

Name Environment Configuration
System Environment Ubuntu 22.04
CPU AMD Ryzen 9 5950X
GPU RTX 4060 Ti 16GB
Deep Learning Framework Pytorch 1.13.1
IDE CUDA 11.7
eISSN:
2470-8038
Langue:
Anglais
Périodicité:
4 fois par an
Sujets de la revue:
Computer Sciences, other
RSS Feed de la revue