Evaluation of deep neural network traffic sign recognition system with focus on data augmentation

Abstract

Purpose:

To develop and evaluate an embedded Traffic Sign Recognition (TSR) system for electromobility applications, emphasizing the role of realistic data augmentation in improving model generalization.

Methods:

A two-stage pipeline was deployed on a Raspberry Pi with Intel Movidius NCS2. It utilizes a fine-tuned SSD MobileNetV2 detector and a lightweight CNN classifier. The model was trained on 60,000 European traffic signs using context-aware augmentations including brightness variation, Gaussian blur, and contrast adjustment.

Results:

The detection network achieved a mean average precision of 0.93 at IoU=0.5. The classification module exceeded 95% accuracy in real-world scenarios. The integrated system sustained real-time operation at 12–15 fps, maintaining high accuracy within the strict thermal and power constraints of the embedded platform.

Conclusions:

Context-aware data augmentation significantly mitigates overfitting and improves robustness against environmental domain shifts. The proposed architecture ensures reliable TSR performance suitable for automotive mechatronics and advanced driver assistance systems (ADAS).