A New Tool for Lane Line Annotation, a New Dataset and a Comparison of Existing Deep Learning-based Lane Line Detection Models

In recent years, the number of autonomous vehicles on the road has been steadily increasing, along with the level of autonomy these vehicles can achieve. The automotive industry is experiencing rapid development, with a strong focus on automating as many tasks as possible that were traditionally performed by the driver. The development of autonomous vehicles and Advanced Driver-Assistance Systems (ADAS) is advancing rapidly, and the scope of their capabilities is expanding. One of the key aspects of ADASs is lane line detection (LLD). This paper presents a new tool for manual lane marking annotation, a critical step in creating high-quality datasets needed for training machine learning-based models. The tool enables intuitive annotation through a graphical interface, project creation, and export of annotations in various formats. Since existing datasets often differ from real-world road conditions, a new custom dataset was created using the new tool, featuring recordings from the roads of the city of Osijek, Croatia. Additionally, the paper includes the performance analysis of LLD methods based on deep learning, with a particular focus on Spatial CNN (SCNN) models, evaluated on both existing and newly created dataset. The results highlight the importance of adapting models to real driving conditions and the potential of new datasets to improve the performance.