Novel automatic traffic sign classification system using a semi-supervised approach

Abstract

Automatic traffic sign classification plays a crucial role in identifying relevant signage that contributes to moving safely autonomous vehicles. Its application is manifold and could also be extended for driving assistance system solutions to help protect the driver and prevent automobile accidents. In this paper we develop an automatic traffic sign classification system using semi-supervised learning, a type of machine learning. The semi-supervised learning lies at the intersection between supervised and unsupervised learning by the fact that the training dataset contains both labeled and unlabeled data. It is a well-known problem that in practice collecting large amounts of labeled samples to train deep learning classifiers is time-consuming and expensive. The semi-supervised learning approach resolves these issues through the use of a partially labeled training dataset. In our experiments we used the SimCLR framework: we pretrained an encoder using the contrastive learning technique on a large set of unlabeled images, and we fine-tuned the encoder on the labeled images. The core idea behind contrastive learning is to learn an embedding space where the objective function minimizes the distance between representations of similar images and maximizes the distance between representations of different images. The dataset we used for training the traffic sign classifier is the German Traffic Sign Recognition Benchmark (GTSRB) dataset which contains 43 traffic sign classes with unbalanced class frequencies. Our work proposes the usage of a novel technique for performing traffic sign classification, semi-supervised learning using the SimCLR framework.