A Novel Object Detection Algorithm Combined YOLOv11 with Dual-Encoder Feature Aggregation.

Area of Science:

• Computer Vision
• Deep Learning
• Robotics

Background:

• Unimodal visual detection struggles in complex environments like low illumination, occlusion, and texture-sparse settings.
• Existing methods often lack robustness and generalization capabilities in diverse and challenging scenarios.

Purpose of the Study:

• To propose an improved YOLOv11-based dual-branch RGB-D fusion framework to overcome limitations of unimodal visual detection.
• To enhance object detection performance in complex scenarios by integrating RGB and depth information.
• To validate the framework's effectiveness and generalization across multiple benchmark datasets and configurations.

Main Methods:

• A symmetric dual-branch architecture processing RGB images and depth maps in parallel.
• Integration of a Dual-Encoder Cross-Attention (DECA) module for cross-modal feature weighting.
• Implementation of a Dual-Encoder Feature Aggregation (DEPA) module for hierarchical fusion.
• Multi-stage evaluation strategy using M³FD and VOC2007 datasets, including RGB-depth, RGB-infrared, and monocular input configurations.

Main Results:

• Achieved mAP50 scores of 82.59% on VOC2007 and 81.14% on M3FD in RGB-infrared mode, outperforming YOLOv11 baseline.
• Attained 77.37% mAP50 with 88.91% precision in RGB-depth on M³FD, demonstrating robustness in geometric-aware detection.
• Ablation studies confirmed the significant contributions of the Dynamic Branch Enhancement (DBE) and Dual-Encoder Attention (DEA) modules.

Conclusions:

• The proposed YOLOv11-based dual-branch RGB-D fusion framework significantly improves object detection accuracy and stability in challenging environments.
• The framework demonstrates strong generalization capabilities across different modalities and datasets.
• The efficient and scalable design offers a promising solution for high-precision spatial perception in autonomous driving and robotics.