Multi-scale target detection for airfield visual navigation of taxiing aircraft

In the context of visual navigation-based civil aircraft assistance systems, the scene target scales captured by onboard cameras exhibit significant variations, and the computational power of airborne platforms is limited. To promptly and accurately identify threat targets in airborne scenic images,...

Full description

Saved in:
Bibliographic Details
Published in2024 4th International Conference on Neural Networks, Information and Communication (NNICE) pp. 749 - 753
Main Authors Luo, Qijun, Chen, Jiagang, Zhang, Xuerui, Zhang, Tao
Format Conference Proceeding
LanguageEnglish
Published IEEE 19.01.2024
Subjects
Aircraft navigation
coordinate attention mechanism
dilation convolution
feature fusion
Measurement
multi-scale target detection
Neck
Reliability
Safety
Visualization
YOLO
Online AccessGet full text

Cover

More Information
Summary:In the context of visual navigation-based civil aircraft assistance systems, the scene target scales captured by onboard cameras exhibit significant variations, and the computational power of airborne platforms is limited. To promptly and accurately identify threat targets in airborne scenic images, a lightweight multi-scale object detection algorithm based on improved YOLOv5s is proposed. Firstly, on the foundation of the Weighted Bidirectional Feature Pyramid Network (BIFPN), a Coordinate Attention (CA) mechanism is integrated to design the CA-BIFPN feature fusion network. This enhances feature representation for small targets and improves the model's learning capability for multi-scale targets. Secondly, a cross-level sub-network module is employed to lightweightly process the neck section, reducing the model's parameter count and thereby improving the overall network's detection speed. Experimental results on a self-constructed scene image dataset demonstrate that the proposed method achieves a detection frame rate of 89FPS, an increase of 5fps. The average precision mean (mAP) on the test set is 70.6%, a 3.4% improvement over the original YOLOv5s. The detection accuracy surpasses typical algorithms such as YOLOv6, YOLOv7, YOLOX, and Faster-RCNN.
DOI:10.1109/NNICE61279.2024.10498484