EfficientIR-Det Towards Efficient and Accurate DETR for UAV Infrared Object Detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Infrared (IR) object detection on unmanned aerial vehicles (UAVs) faces challenges from low signal-to-noise ratios and limited onboard computational power.
• Existing Convolutional Neural Networks (CNNs) lack global context, while Transformers exhibit quadratic complexity, impeding real-time deployment.

Purpose of the Study:

• To develop a lightweight, end-to-end infrared object detector for UAVs that overcomes computational and performance bottlenecks.
• To introduce novel architectural components for efficient feature extraction and global context modeling in resource-constrained environments.

Main Methods:

• Proposed EfficientIR-Det, a holistic detector optimizing backbone, encoder, and sampling.
• Introduced Partial Star Network (PSN) backbone for implicit high-dimensional feature expansion and weak IR signal amplification.
• Developed Hierarchical Mamba (HiMamba) encoder for linear-complexity global enhancement and Adaptive Gated Sampling (AGS) with Hierarchical Sampling Strategy (HSS) for refined cross-scale representation.

Main Results:

• Achieved 88.4% mAP@0.5 on HIT-UAV, surpassing RT-DETR-R18 by 3.3 points.
• Reduced FLOPs by 48.9% and parameters by 44.2% compared to baseline.
• Attained 74.1% mAP@0.5 and 140.8 FPS on the DroneVehicle dataset.

Conclusions:

• EfficientIR-Det offers a promising solution for robust, real-time infrared perception on edge-constrained UAV platforms.
• The proposed methods demonstrate significant improvements in detection accuracy and computational efficiency for UAV-based IR object detection.