A Novel One-Step Small Object Detector for Autonomous Aerial Vehicles.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Detecting small objects in aerial imagery is difficult due to scale variations and non-uniform distribution.
• Autonomous aerial vehicles require efficient and accurate detection methods with limited computational resources.
• Existing Feature Pyramid Network (FPN)-based methods struggle with noise and high computational costs in feature fusion for small objects.

Purpose of the Study:

• To develop a novel one-step generative small object detection (SOD) framework for aerial images.
• To enhance detection accuracy and efficiency for small objects, addressing limitations of current methods.
• To provide a computationally efficient solution for autonomous aerial vehicles.

Main Methods:

• Formulated small object detection as a noise-to-box procedure using a consistency model.
• Leveraged the self-consistency property of a consistency model for one-step inference from Gaussian noise to a single-scale output.
• Employed a denoising sampling strategy to iteratively refine Gaussian distributions for classifying and locating small objects.

Main Results:

• The proposed framework achieved superior performance compared to state-of-the-art methods.
• Demonstrated up to a 5.1% improvement in average precision on small objects ($AP_{S}$) on the DOTA benchmark.
• Successfully evaluated on DOTA, VisDrone, and AAVDT datasets, confirming effectiveness for autonomous aerial vehicles.

Conclusions:

• The novel one-step generative framework offers a significant advancement in small object detection for aerial imagery.
• The approach effectively balances accuracy and efficiency, making it suitable for resource-constrained autonomous aerial vehicles.
• The method shows strong potential for real-world applications in aerial surveillance and autonomous navigation.