Vehicle Detection in Drone Aerial Views Based on Lightweight YOLOv10-IAD

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Unmanned Aerial Vehicle (UAV)-based vehicle detection is crucial for traffic monitoring and management.
• Existing methods struggle with challenges like small targets, dense distributions, and occlusions in aerial imagery.

Purpose of the Study:

• To develop an advanced YOLOv10-based model (YOLOv10-IAD) for robust UAV-based vehicle detection.
• To enhance the detection of small, dense, and occluded vehicles in aerial datasets.

Main Methods:

• Integrated Involution convolution in the backbone for improved small target perception.
• Incorporated Attention and Convolution Mixed (ACmix) in the neck for fused feature representation.
• Utilized Dynamic Head (DyHead) for attention-based feature recalibration to handle occlusions.

Main Results:

• Achieved significant improvements in mean Average Precision (mAP50) and recall on VisDrone2019 and UAVDT datasets.
• Demonstrated a favorable trade-off between detection accuracy and computational efficiency compared to other YOLO models.
• YOLOv10-IAD showed a 3.7% and 3.5% increase in mAP50 on the respective datasets.

Conclusions:

• YOLOv10-IAD effectively addresses key challenges in UAV-based vehicle detection.
• The proposed model offers superior performance with minimal increases in parameters and computational cost.
• Suitable for real-time vehicle detection applications on UAV onboard hardware.