Vehicle detection and classification using an ensemble of EfficientDet and YOLOv8

Area of Science:

• Computer Science
• Artificial Intelligence
• Electrical Engineering

Background:

• Traffic management faces challenges due to increasing vehicle numbers and limitations of traditional detection methods.
• Diverse vehicle characteristics (shape, color, texture) complicate accurate identification.
• Existing deep learning models may struggle with varied imaging conditions and class imbalances.

Purpose of the Study:

• To develop an improved vehicle detection and classification system for intelligent traffic management.
• To leverage ensemble deep learning for enhanced accuracy and robustness.
• To evaluate the performance of an ensemble model using both thermal and RGB imagery.

Main Methods:

• An ensemble method combining EfficientDet and YOLOv8 deep learning models was proposed.
• The Forward-Looking Infrared (FLIR) dataset, containing thermal and RGB images, was utilized.
• Data augmentation techniques were applied to improve model performance and address class imbalances.

Main Results:

• The ensemble model achieved a 95.5% mean average precision (mAP) on thermal images, surpassing individual models.
• On thermal images, the ensemble model recorded an average recall (AR) of 0.93 and an optimal localization recall precision (oLRP) of 0.08.
• For RGB images, the ensemble model achieved 93.1% mAP, 0.91 AR, and 0.10 oLRP.

Conclusions:

• The proposed ensemble approach significantly enhances vehicle detection and classification accuracy.
• Integration of thermal imaging improves detection under various lighting conditions, ensuring system robustness.
• The developed system offers a robust solution for real-world intelligent traffic management applications.