An improved UAV target detection algorithm based on ASFF-YOLOv5s.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Object detection in drone-captured scenarios faces challenges like high altitude, scale variation, and dense occlusion, demanding real-time performance.
• Existing algorithms struggle with accurately detecting small targets in complex aerial environments.

Purpose of the Study:

• To develop a real-time Unmanned Aerial Vehicle (UAV) small target detection algorithm with improved accuracy and efficiency.
• To enhance feature extraction and fusion capabilities for small objects in aerial imagery.

Main Methods:

• An improved ASFF-YOLOv5s algorithm was proposed, incorporating multi-scale feature fusion and enhanced Adaptively Spatial Feature Fusion (ASFF).
• K-means clustering was adapted to generate optimal anchor frames for the VisDrone2021 dataset.
• Convolutional Block Attention Module (CBAM) was integrated to refine feature capture, and SIoU loss function was employed for improved convergence and accuracy.

Main Results:

• The proposed model achieved 70.4 FPS with a precision of 32.55%, F1-score of 39.62%, and mAP of 38.03% on the VisDrone2021 dataset.
• Significant improvements of 2.77% (precision), 3.98% (F1-score), and 5.1% (mAP) were observed compared to the original YOLOv5s algorithm.
• The model demonstrated effective detection of small targets in diverse and challenging aerial environments.

Conclusions:

• The developed algorithm provides an effective solution for real-time small target detection in UAV aerial photography, even in complex scenes.
• The method shows potential for extension to other applications such as urban security surveillance for detecting pedestrians and vehicles.