Validation of Machine Learning-Based Lane Lines Detection Methods Using Different Datasets

Abstract

Advanced Driver Assistance System (ADAS) uses various sensors that exist in the vehicle to collect information about the vehicle and its surroundings. One of the most commonly used ADASs is the one designed to recognize driving lane lines. In this paper, analysis, and validation of freely available state-of-the-art deep learning-based lane lines detection (LLD) methods are done using three different well-known datasets: CULane, TuSimple, and LLAMAS. To perform the validation, as part of this research, two converters of lane line label formats were made. Converters allow testing models that have been previously trained on a dataset that uses different lane line label formats so that their performances can be verified on other datasets with different lane line format types. The experimental results show that the adaptability of the model to the change of the dataset depends on the architecture used as the basis of the model for the LLD, and the similarity of the training set of the initially used dataset and the new, i.e. test dataset. All tested models showed shortcomings when tested on images from datasets on which they were not trained, which confirms that there is still no model that we can confidently say achieves high performance in all possible scenarios.