DATASET CONSTRUCTION AND IMAGE PERSPECTIVE ANALYSIS FOR LANE DETECTION UNDER VIETNAMESE ROAD CONDITIONS

Abstract

This article presents the construction of a dataset for lane detection research under Vietnamese road conditions along with an evaluation of some of the best state-of-the-art lane detection models at the moment, as well as an analysis of image perspective effects on detection performance. A custom dataset is collected from real-world road scenes in Vietnam, covering diverse traffic scenarios and lane marking conditions. Multiple pretrained state-of-the-art lane detection models are then evaluated on the proposed dataset using precision, recall, and F1-score metrics. To study perspective sensitivity, perspective modification was applied to the input images, shifting the vanishing-line (road horizon) from 20% to 80% of the image’s height. The results indicate a clear dependence on image perspective, with the best overall performance achieved (F1-score peaks at 73.4%) when the vanishing-line position is 37% of the image height. The dataset, perspective analysis, and benchmarking results together provide practical guidance for dataset design, perspective normalization, and the deployment of pretrained lane detection models in real-world Vietnamese traffic environments.