A traffic and road signal recognition method through deep learning

Abstract

Before deep learning became popular, some researchers used traditional computer vision algorithms, including support vector machines, decision trees, and random forests to deal with traffic and road signal recognition problems. These methods often require manual design of features and may not perform well when dealing with complex scenarios and changing conditions. Therefore, traffic and road signal recognition have witnessed a transformative shift with the advent of deep learning technologies. Convolutional neural networks (CNNs) developed from deep learning has shown prominent capabilities in accurately detecting and interpreting traffic signs and signals from images and videos. In this project, the whole method is mainly based on two parts, which refers to the process of signal-capturing and the detecting process. The CNN model used in the latter part consists of various function layers, including max-pooling layer, linear layer, 2D convolution layer and batch- normalization layer to modify accuracy. This model is trained and tested with Chinese Traffic Sign Dataset, with 58 categories of over 6000 pictures. With the assistance of OpenCV, the model may detect real-time traffic signals and convey messages to vehicles. After adjusting parameters in the CNN model, high probability tags are obtained during the process, with both accuracy and average loss labeled after each epoch. The results demonstrate that the proposed model achieved the testing accuracy of over 95 percent, after undergoing 200 epochs.