Study on lightweight strategies for L-YOLO algorithm in road object detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Autonomous Systems

Background:

• Urban traffic complexity necessitates efficient object detection for autonomous driving and traffic management.
• Traditional object detection algorithms face limitations due to large parameter sizes and high computational costs, hindering deployment in resource-constrained environments.

Purpose of the Study:

• To develop a lightweight and efficient road object detection algorithm, L-YOLO, based on YOLOv8s.
• To enhance feature extraction, small object detection, model robustness, and computational efficiency.

Main Methods:

• Replaced YOLOv8s backbone with L-HGNetV2 for improved feature extraction and fusion.
• Introduced a small object detection layer with CStar modules to enhance small vehicle feature capture.
• Implemented FPIoU2 loss function for improved model robustness.
• Applied layer adaptive magnitude-based model pruning (LAMP) to reduce parameters and computational load.

Main Results:

• L-YOLO achieved a mAP50 of 93.8% on the KITTI dataset, a 2.5% improvement over YOLOv8s.
• Reduced model parameters from 11.12 M to 3.58 M.
• Decreased computational load from 28.4 GFLOPs to 14.2 GFLOPs.

Conclusions:

• L-YOLO offers a significant improvement in efficiency and accuracy for road object detection compared to YOLOv8s.
• The proposed lightweight algorithm is suitable for resource-constrained environments in autonomous driving and intelligent traffic management.