A Survey of Data Augmentation Techniques for Traffic Visual Elements.

Area of Science:

• Computer Science
• Artificial Intelligence
• Robotics

Background:

• Autonomous driving relies heavily on visual perception of traffic elements (signs, lights, pedestrians).
• Current datasets for autonomous driving often lack diversity and suffer from class imbalance, hindering model robustness.
• A systematic review of dataset augmentation specifically for traffic visual elements is needed.

Purpose of the Study:

• To systematically analyze enhancement techniques for transportation datasets in autonomous driving.
• To establish a classification framework for autonomous driving scenarios and assess augmentation's impact on detection and classification.
• To offer practical guidance for improving autonomous driving datasets in research and industry.

Main Methods:

• Analysis of four augmentation approaches: image transformation, Generative Adversarial Networks (GANs), diffusion models, and composite methods.
• Review of nearly 40 traffic-related datasets and 10 evaluation metrics for benchmarking.
• Assessment of performance gains from augmentation on object detection and classification tasks.

Main Results:

• Dataset augmentation significantly improves the robustness of autonomous driving models, especially under challenging conditions.
• Hybrid augmentation methods generally yield the best performance improvements.
• Key challenges include high computational costs, unstable GAN training, and insufficient data for rare scenarios.

Conclusions:

• Augmentation is crucial for enhancing the reliability of visual perception in autonomous driving systems.
• Hybrid and advanced generative methods show strong potential but require further optimization.
• Future research should focus on efficient models, richer semantic context, specialized datasets, and scalable augmentation strategies.