Lane detection in miniature environments using image segmentation


Lane detection is an important task when it comes to de- veloping advanced driver assistance systems (ADAS) and autonomous vehicles. While much research is dedicated to real world vehicles, we want to try a machine learning based lane detection approach in a 1:87 miniature environment. With a fixed network architecture, we experi- ment with different training methods to learn how both existing real world data and specialized training data for our environment affects lane detection performance. While only using real world training data granted usable results, using a low number of specialized training images greatly improved performance. Yet, further experiments are necessary for reliable lane detection.