YOLO-PEL: The Efficient and Lightweight Vehicle Detection Method Based on YOLO Algorithm.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Vehicle detection is crucial for intelligent transportation systems.
• Fixed camera systems often face trade-offs between detection accuracy, cost, and real-time performance.
• Existing models may struggle with feature fusion and generalization across scales.

Purpose of the Study:

• To develop an efficient and accurate vehicle detection model for resource-constrained, real-time applications.
• To enhance the YOLOv8n model for improved feature adaptability and multi-scale integration.
• To address challenges in vehicle detection precision caused by extreme samples.

Main Methods:

• Introduced the C2F-PPA module for enhanced feature fusion in YOLOv8n.
• Proposed ELA-FPN for refined multi-scale feature fusion and generalization.
• Incorporated the Wise-IoUv3 loss function to improve detection accuracy.
• Trained and evaluated the model on COCO-Vehicle and VisDrone2019 datasets.

Main Results:

• YOLOv8-PEL achieved 66.9% mAP@0.5 on the COCO-Vehicle dataset.
• The model boasts 2.23 M parameters, 7.0 GFLOPs, 4.5 MB size, and 176.8 FPS inference speed.
• Achieved significant reductions in parameters (25%), GFLOPs (13%), and model size (25%) compared to YOLOv8n.
• Demonstrated excellent computational efficiency and generalization capability.

Conclusions:

• YOLOv8-PEL offers a superior balance of detection accuracy and computational efficiency.
• The model is highly suitable for real-time and resource-constrained vehicle detection scenarios.
• The proposed enhancements effectively improve feature fusion and mitigate gradient issues for precise detection.