DTC-YOLO: Multimodal Object Detection via Depth-Texture Coupling and Dynamic Gating Optimization.

Area of Science:

• Computer Vision
• Machine Learning
• Sensor Fusion

Background:

• Single-modality sensors face limitations due to physical properties and data types.
• Existing object detection models struggle with scale variations in complex environments.

Purpose of the Study:

• To introduce DTC-YOLO, a novel depth-texture coupled multimodal detection framework.
• To enhance object detection accuracy by effectively integrating RGB and LiDAR data.

Main Methods:

• Developed a depth-color mapping and weighted fusion strategy for RGB-LiDAR integration.
• Introduced ADF³-Net, a feature fusion network with adaptive, hierarchical, and decoupled processing.
• Implemented an Adown Module for efficient downsampling, separating high-frequency details and low-frequency semantics.

Main Results:

• DTC-YOLO achieved significant improvements: +3.50% mAP50, +3.40% mAP50-95, and +3.46% precision.
• The framework demonstrated enhanced detection of extremely large and small objects.
• Reduced Giga Floating-point Operations Per Second (GFLOPs) by 10.53% while maintaining performance.

Conclusions:

• DTC-YOLO effectively mitigates scale-related accuracy discrepancies common in vision-only models.
• The proposed depth-texture coupling mechanism offers a robust solution for multimodal object detection.
• This framework shows promise for improving autonomous driving systems in complex traffic scenarios.