An algorithm for road target detection of autonomous vehicles based on improved YOLOv8

Area of Science:

• Computer Vision
• Artificial Intelligence
• Autonomous Systems

Background:

• Autonomous driving technology relies heavily on accurate road target identification by onboard cameras.
• Complex traffic environments lead to challenges in detecting near and far road targets.
• Existing algorithms struggle with false detections of overlapping objects and missed detections of distant objects.

Purpose of the Study:

• To develop an improved road target detection algorithm for autonomous vehicles.
• To address false detection of near-range overlapping targets and missed detection of long-range targets.
• To enhance the overall accuracy and reliability of road object recognition.

Main Methods:

• Proposed YOLOv8-RTDAV algorithm based on YOLOv8n.
• Introduced C2f-EFB module with efficient channel attention (ECA) for close-range targets.
• Added a P2 small target detection layer for long-range targets.
• Replaced SPPF with SPPELAN module and used DySample for improved feature capture.
• Enhanced the loss function from CIOU to EIOU.

Main Results:

• YOLOv8-RTDAV demonstrated improved accuracy on KITTI and TT100K datasets compared to YOLOv8n.
• Successfully improved recognition of close-range overlapping road targets.
• Increased the recognition rate of small targets at long distances.

Conclusions:

• The proposed YOLOv8-RTDAV algorithm effectively enhances road target detection for autonomous vehicles.
• The improvements address critical challenges in real-world driving scenarios.
• This contributes to safer and more reliable autonomous driving systems.