Vehicle detection in drone aerial views based on lightweight OSD-YOLOv10.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Existing vehicle detection models for UAV imagery face challenges with low performance, small target feature extraction, and large parameter sizes.
• Efficient and accurate object detection is crucial for autonomous systems and surveillance using Unmanned Aerial Vehicles (UAVs).

Purpose of the Study:

• To develop an optimized YOLOv10-based algorithm (OSD-YOLOv10) for enhanced vehicle detection in UAV aerial imagery.
• To address limitations in small target detection and reduce computational complexity for deployment on resource-constrained UAV hardware.

Main Methods:

• Introduced OCRConv module via online convolutional reparameterization and a lightweight SPCC structure replacing the C2f module.
• Integrated a dual-layer feed-forward hybrid attention module and a dual small-target detection layer combining shallow features.
• Implemented the DySample dynamic upsampling module for improved feature fusion in the neck network.

Main Results:

• OSD-YOLOv10 achieved a 40.7% reduction in parameters and a 3.6% decrease in floating-point operations.
• Demonstrated improvements in accuracy (+1.3%) and mean average precision (+1.6%) on VisDrone-DET2019 and UAVDT datasets.
• Outperformed other YOLO series and lightweight models in detection accuracy and computational complexity.

Conclusions:

• OSD-YOLOv10 offers an optimal balance between high accuracy and low resource consumption for UAV-based vehicle detection.
• The proposed algorithm is highly suitable for deployment on UAV onboard hardware, enabling efficient real-time target detection.
• The advancements contribute to more capable and efficient autonomous systems utilizing aerial imagery.