YOLO11s-UAV: An Advanced Algorithm for Small Object Detection in UAV Aerial Imagery.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Unmanned aerial vehicles (UAVs) are increasingly vital for applications like agriculture and search and rescue.
• Detecting small objects in aerial imagery presents challenges due to sparse pixels, complex backgrounds, and limited onboard computational power.
• Existing object detection algorithms struggle with the unique demands of aerial surveillance.

Purpose of the Study:

• To develop an improved UAV-based small object detection algorithm for enhanced performance in aerial imagery.
• To address challenges of small object detection, complex backgrounds, and computational constraints in UAV applications.
• To create a more efficient and accurate model for real-time object detection on UAV platforms.

Main Methods:

• Introduced Content-Aware Reassembly and Interaction Feature Pyramid Network (CARIFPN) for superior small object feature detection and reduced network complexity.
• Implemented Space-to-Depth for Dilation-wise Residual Convolution (S2DResConv) in the backbone to prevent information loss and capture multi-scale context.
• Integrated Flexible SimAM (FlexSimAM), a parameter-free attention module, to efficiently enhance small object features in complex aerial scenes.

Main Results:

• Achieved a 7.8% improvement in mAP@0.5 on the VisDrone-DET2019 validation set (46.0%) and a 5.9% increase on the test set (37.3%) compared to the YOLO11s baseline.
• Reduced model parameters by 55.3% and demonstrated significant improvements on TinyPerson (7.2%) and UAVDT-DET (3.0%) datasets.
• Attained 33 FPS on the NVIDIA Jetson Orin NX SUPER, confirming real-time onboard processing feasibility with a 21.4% reduction in processing time.

Conclusions:

• The proposed YOLO11s-UAV algorithm effectively enhances small object detection in aerial imagery.
• The novel CARIFPN, S2DResConv, and FlexSimAM modules contribute to improved accuracy and efficiency.
• The model's performance and reduced complexity make it suitable for real-time onboard UAV applications.