Neural Network Architectures for Lane Detection and Road Segmentation in Self-Driving Cars

Abstract

Road and lane detection are essential components for the environmental perception of Advanced Driver Assistance Systems (ADAS) and self-driving cars (SDC) [1]. It predicts the map specifications that will be used for localization and planning algorithms. As well, it is critical when it comes to path planning in ADAS models. While there are many well-developed methods for lane detection and segmentation, the most important challenge is the generalization problem related to the dataset differences in real-world applications. Classes that are not found in the pre-trained model, ad-hoc tuning instead of training on real-world instances, limited generalization potential due to the small batch sizes, handcrafted feature and synthetic data usage, ad-hoc region proposal methods will lead to a non-robust network on real-world application data even though the segmentation network in question could be successful at the image dataset [2]. This is why networks overfit to dataset related to the problem of encoder-decoder architectures. However, a few studies have been done to explicitly investigate the generalization potential of encoder-decoder architectures. There are only several quantification reports with large generalization benchmark dataset analysis to advantage encoder-decoder architectures in the area of Object Detection and Segmentation.