Recent Real-Time Aerial Object Detection Approaches, Performance, Optimization, and Efficient Design Trends for...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Growing demand for real-time aerial perception necessitates efficient onboard object detection.
• Edge devices and UAVs have severe constraints on computation, memory, and power.

Purpose of the Study:

• To survey real-time aerial object detection algorithms for constrained onboard environments.
• To analyze datasets, challenges, and optimization techniques for aerial object detection.

Main Methods:

• Review of aerial and remote-sensing datasets and their associated challenges (small objects, scale variation, complex backgrounds).
• Categorization of recent studies on lightweight backbones, feature representation, multi-scale fusion, and optimized Transformers.
• Highlighting hardware-aware optimization (quantization, pruning, TensorRT) and Neural Architecture Search (NAS).

Main Results:

• Identification of key challenges in aerial object detection.
• Overview of algorithmic advancements for lightweight and efficient detectors.
• Discussion of model compression and deployment optimization for real-time performance.

Conclusions:

• A unified perspective on designing next-generation aerial object detectors is presented.
• Achieving both high accuracy and real-time performance on UAVs is feasible through integrated strategies.
• Future research directions include automated NAS tailored for UAV constraints and adapting large pretrained models.