RD-YOLO: An Effective and Efficient Object Detector for Roadside Perception System.

Area of Science:

• Intelligent Transportation Systems
• Computer Vision
• Deep Learning for Autonomous Driving

Background:

• Vehicle-road cooperation is a key area in intelligent transportation, yet roadside perception remains less studied than vehicle perception.
• Challenges in roadside perception include drastic variations in object scale due to sensor height and difficulties in distinguishing overlapping or occluded objects in complex environments.

Purpose of the Study:

• To address the limitations in roadside object detection, this study proposes RD-YOLO, an improved object detection algorithm.
• The goal is to enhance the detection of small roadside objects and improve the network's adaptability to varying object scales and occlusions.

Main Methods:

• RD-YOLO is built upon YOLOv5s, featuring a reconstructed feature fusion layer for better small target detection and a generalized feature pyramid network (GFPN) for improved scale adaptability.
• A coordinate attention (CA) mechanism is integrated to focus on relevant regions in dense object scenarios, and Focal-EIOU Loss is used to optimize bounding box regression speed and anchor box positioning accuracy.

Main Results:

• RD-YOLO demonstrated significant improvements, increasing mean average precision (mAP) by 5.5% on the Rope3D dataset and 2.9% on the UA-DETRAC dataset compared to YOLOv5s.
• The modified feature fusion layer reduced RD-YOLO's weight by 55.9% with minimal impact on detection speed, achieving over 71.9 frames per second (FPS).

Conclusions:

• RD-YOLO effectively tackles challenges in roadside perception, offering a more accurate and efficient solution for intelligent transportation systems.
• The algorithm achieves real-time detection capabilities with superior accuracy compared to existing methods at similar speeds, making it suitable for practical applications.