Real-Time Embedded Traffic Sign Recognition Using Efficient Convolutional Neural Network

Bangquan, Xie; Xiong, Weng Xiao · 2019 · OpenAlex-citations

DOI: 10.1109/access.2019.2912311

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of deploying Traffic Sign Recognition (TSR) systems on computationally limited embedded platforms for autonomous driving. TSR comprises two sub-tasks: Traffic Sign Classification (TSC) and Traffic Sign Detection (TSD). While deep learning models have improved accuracy, their increasing depth and parameter counts hinder real-time performance. The authors aim to develop efficient Convolutional Neural Networks (CNNs) that balance accuracy, generalization, and speed. To this end, they propose two novel architectures: ENet for classification and EmdNet for detection. The methodology involves rigorous data analysis and iterative network construction. For TSC, the authors utilized the German Traffic Sign Recognition Benchmark (GTSRB). They employed data mining to analyze dataset regularities, discovering that consecutive images in the training set were highly similar due to continuous camera capture. This insight informed their validation set division and data augmentation strategies, which included creating "Equaled" and "Enlarged" datasets using rotation, reflection, and flipping to address class imbalance. The ENet architecture was developed through an eight-step experimental process, evaluating kernel sizes, grayscale conversion, normalization, dropout, mixed kernels, shortcuts, and Depthwise Separable Convolutions (DSC). For TSD, the authors used the LISA US Traffic Sign Dataset and designed EmdNet based on a modified Single Shot MultiBox Detector (SSD) framework, incorporating multiscale operations and DSC to handle real-world challenges like occlusion and illumination changes. The results demonstrate significant efficiency gains without compromising accuracy. ENet-V1, the most accurate variant, achieved 98.6% accuracy on GTSRB with only 0.9 million parameters, which is 1/15th the parameters of state-of-the-art methods. ENet-V2, optimized for speed using DSC, processed samples in 0.62 ms. EmdNet, designed for detection, contained only 6.3 million parameters, comparable to MobileNet, and successfully handled partial occlusion, illumination variations, and cluttered backgrounds. The study confirms that DSC and shortcut connections effectively reduce computational load while maintaining robust feature extraction. The significance of this work lies in providing lightweight, high-performance models suitable for real-time embedded systems in autonomous vehicles. By reducing parameter counts by orders of magnitude compared to traditional deep networks like VGG or ResNet, ENet and EmdNet enable efficient deployment on hardware with limited processing power. The paper highlights that efficiency can be achieved through architectural innovations like DSC and data-driven strategies, rather than solely relying on model compression techniques. This contributes to the broader field of efficient deep learning by offering practical solutions for perception tasks in safety-critical applications.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• sign visibility legibility