Research on Improved Algorithms for Cone Bucket Detection in Formula Unmanned Competition

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Object detection models like YOLOv8 face challenges in complex scenarios, including large parameter counts and computational demands.
• Efficient detection of race cones and buckets is crucial for autonomous vehicle competitions, such as the Formula Unmanned Competition.

Purpose of the Study:

• To develop a lightweight object detection model based on YOLOv8 for improved efficiency in detecting race cones and buckets.
• To address limitations in complex structure, parameter redundancy, and computation affecting detection performance.

Main Methods:

• Proposed a lightweight detection model by enhancing the backbone network (using YOLOv9's ADown module), neck network (replacing YOLOv8 C2f's fusion module with FasterNet's FasterBlock), and detection head.
• Integrated knowledge distillation to further optimize detection performance.
• Utilized the FSACOCO dataset for experimental validation.

Main Results:

• The improved model achieved an accuracy of 92.7%, recall of 84.6%, and average precision of 91% on the FSACOCO dataset.
• Demonstrated a 2.7% increase in recall and 1.2% increase in average precision compared to the original YOLOv8n model.
• Reduced model memory by 50% and computation by 51%, while significantly decreasing misdetections and ensuring high detection speed.

Conclusions:

• The developed lightweight YOLOv8 model effectively enhances the detection of race cones and buckets, meeting the stringent requirements for deployment on tiny devices in autonomous racing.
• The proposed improvements offer a viable solution for object detection in complex environments and can be extended to other small target detection tasks.