A Lightweight Model for Traffic Sign Recognition Based on Attention Mechanism

Abstract

Demand for reliable systems that can quickly and accurately recognize textual traffic signals is growing as autonomous vehicles become more widespread around the world. Because they are more complicated and have fewer datasets than symbolic traffic signals, textual traffic signals particularly receive less attention than symbolic ones. In this research, we build a lightweight CNN model with layers for attention mechanism that can rapidly and accurately classify 44 different traffic signs, including our new class related to textual traffic signs (panels) and 43 classes from the standard German Traffic Sign Recognition (GTSRB) dataset. Our model outperformed the state-of-the-art models with remarkable results, obtaining test accuracy of 98.7% and training accuracy of 99.89% with a substantially smaller number of parameters.