Road Traffic Classification from Nighttime Videos Using the Multihead Self-Attention Vision Transformer Model and the SVM

Abstract

Intelligent transport systems (ITSs) have emerged as a groundbreaking solution to address the challenges associated with road traffic, which are enhancing road utilization efficiency, providing convenient and safe transportation, and reducing energy consumption. ITS leverages advanced technologies to collect, store, and deliver real-time road traffic information, enabling intelligent decision-making and optimizing various aspects of transportation systems. As a contribution in this matter, we propose in this paper a novel efficient macroscopic approach, based on the multihead self-attention vision transformer (MSViT), for categorizing road traffic congestion, from nighttime videos, into three classes: light, medium, and heavy. To assess the performance of our approach, we conducted experiments using the nighttime UCSD (University of California San Diego) dataset, which includes various weather conditions (clear, overcast, and rainy) and traffic scenarios (light, medium, and heavy). The classification accuracy reached a high level of 94.24%. By incorporating a support vector machine (SVM) classifier into this method, we managed to enhance this accuracy to the outstanding level of 98.92%, thus outperforming the existing state-of-the-art methods that were evaluated using the same UCSD dataset, furthermore, the execution time was optimized.