Real-Time Vehicle Classification and Tracking Using a Transfer Learning-Improved Deep Learning Network.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Intelligent Transport Systems

Background:

• Accurate vehicle classification and tracking are crucial for intelligent transport systems (ITSs) and location-based planning.
• Deep learning (DL) and computer vision methods face challenges with large datasets, domain shift, and real-time multi-vehicle tracking integration.

Purpose of the Study:

• To develop a robust system for real-time vehicle classification and tracking.
• To overcome limitations in training data availability and domain adaptation for DL models.
• To integrate a DL-based classifier with a novel multi-vehicle tracking algorithm.

Main Methods:

• Created a 30,000-sample vehicle dataset with seven classes.
• Applied transfer learning and fine-tuning to state-of-the-art YOLO networks to address domain shift.
• Developed a real-time multi-vehicle tracking algorithm for per-lane counting, classification, and speed estimation.

Main Results:

• Fine-tuning doubled classification accuracy from 30% to 71%.
• The YOLOv5-large network coupled with the tracking algorithm achieved 95% accuracy.
• YOLOv5-large offered an optimal balance of accuracy, low loss (0.033), and smaller model size (91.6 MB) compared to other YOLO networks.

Conclusions:

• The proposed approach effectively addresses key challenges in real-time vehicle analysis for ITSs.
• Fine-tuned DL models significantly enhance vehicle classification and tracking performance.
• The integration of YOLOv5-large with the custom tracking algorithm provides a practical solution for intelligent transport planning and spatial information management.