Lane Detection and Tracking Datasets: Efficient Investigation and New Measurement by a Novel "Dataset Scenario Detector" Application

Hossein Maghsoumi, Nasser Masoumi, Babak Nadjar Araabi

Abstract: Autonomous driving research is critically dependent on extensive amounts of real-world data to develop, test, and evaluate systems before deployment in public roadways. At different levels of autonomous driving, one of the most significant and fundamental parts for understanding the environment is the lane detection and tracking system. However, there is no fair comparison between algorithms because of the lack of publicly available datasets, standard evaluation procedures, and uniform metrics in past studies. This article presents a comprehensive study of publicly available datasets to support the research and development of algorithms in lane detection and tracking systems. Depending on the scenes involved, each lane dataset challenges the accuracy and robustness key performance indicators (KPIs) of the lane detection algorithms to varying degrees. For the first time, we present a new measurement on the lane datasets that displays the number of challenging scenarios in the KPIs parameters. For this purpose, we designed and implemented the effective “dataset scenario detector” application, which allows the analysis of large datasets and has Excel outputs.