Lightweight Vehicle Detection Based on Mamba_ViT.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Vehicle detection is crucial for intelligent traffic management and autonomous driving.
• Current deep learning methods using convolutional neural networks (CNNs) have limitations in fully capturing complex traffic scenarios.
• Relying on single feature extraction units hinders comprehensive vehicle information understanding.

Purpose of the Study:

• To develop a lightweight and effective vehicle detection algorithm.
• To improve the accuracy and efficiency of vehicle detection in real-world traffic scenarios.
• To address the limitations of single feature extraction in current deep learning models.

Main Methods:

• Proposed a novel feature extraction architecture, Mamba_ViT, which independently processes shallow and deep features for richer contextual representation.
• Introduced a multi-scale feature fusion mechanism to integrate diverse features effectively.
• Developed the Mamba_ViT_YOLO algorithm by combining the Mamba_ViT architecture with a YOLO framework.

Main Results:

• The Mamba_ViT_YOLO algorithm demonstrated improved performance on the UA-DETRAC dataset.
• Achieved a 3.2% increase in mAP@50 compared to the YOLOv8 algorithm.
• The proposed model utilizes only 60% of the parameters compared to YOLOv8, indicating enhanced efficiency.

Conclusions:

• The Mamba_ViT_YOLO algorithm offers a significant advancement in vehicle detection.
• The novel Mamba_ViT architecture and feature fusion mechanism contribute to more comprehensive feature extraction.
• This lightweight model provides a promising solution for real-time intelligent traffic management and autonomous driving systems.