Classification of Road Features Using Transfer Learning and Convolutional Neural Network (CNN)

Abstract

Efficient and accurate classification of road features, such as crosswalks, intersections, overpasses, and roundabouts, is crucial for enhancing road safety and optimizing traffic management. In this study, we propose a classification approach that utilizes the power of transfer learning and convolutional neural networks (CNNs) to address the road feature classification problem. By leveraging advancements in deep learning and employing state-of-the-art CNN architectures, the proposed system aims to achieve robust and realtime classification of road features. The dataset contained 7616 images of roundabouts, crosswalks, overpasses, and intersections from the MLRSNet dataset and manually extracted satellite images from Malaysia using Google Earth Pro. After that, we merged this dataset. We designed a CNN architecture that consists of 24 convolution layers and eight fully connected layers. Transfer learning models such as ResNet50, MobileNetV2, VGG19 and InceptionV3 were also explored for road feature classification. The best-performing model during the validation phase is InceptionV3, with an accuracy of 98.9777%, whereas the best-performing model during the test phase is ResNet50 and VGG-19 models, with an accuracy of 98.7132%. The proposed CNN model got 95.1208% and 94.4852% accuracy during the validation and test stage. From the evaluation, the best-performing models for road feature classification are ResNet50 and VGG-19, with an accuracy of 98.7132%.