DER-YOLO: A Lightweight Stage-Wise Feature Calibration Network for Onboard Real-Time Small-Object Detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Accurate real-time detection of small traffic objects is crucial for onboard vision-based traffic perception systems.
• Challenges include weak texture, scale variation, occlusion, and limited computational resources in complex traffic scenes.

Purpose of the Study:

• To propose DER-YOLO, a lightweight, small-object-oriented detector designed for complex traffic scenes.
• To enhance small-object representation and localization in onboard traffic perception.

Main Methods:

• DER-YOLO builds upon YOLO11n, incorporating stage-wise feature calibration.
• Key modules include Decoupled Global Context C3k2 (DGC-C3k2), ECA-guided Cross-scale Adaptive Fusion (ECAF), and Refined Large Selective Kernel (RLSK).

Main Results:

• DER-YOLO demonstrated improved detection accuracy on KITTI and BDD100K datasets while maintaining real-time inference.
• Achieved 86.95% mAP@0.5 on KITTI and 55.05% mAP@0.5 on BDD100K, with significant gains in small-object AP.
• Operates with 2.744M parameters, 7.401 GFLOPs, and over 100 FPS.

Conclusions:

• DER-YOLO offers an effective and lightweight solution for real-time small-object detection in onboard traffic perception.
• The proposed stage-wise feature calibration effectively addresses challenges in complex traffic environments.