A Review of Neural Network based Semantic Segmentation for Scene Understanding in Context of the self driving Car

This paper tackles the challenge of scene understanding in context of automated driving. To react properly to the conditions given by the surrounding scene, the car has to understand it’s environment. Further the real time capability of a method solving this task is essential. For scene understanding the car has to detect and classify surrounding objects. For this purpose one can employ a semantic segmentation to assign a class label to every pixel. In this paper we evaluate the the state of the art methods for the semantic segmentation and perform tests on the FCN-8 architecture. Due to Hardware limitations, we train the FCN-8 on a downscaled version of the Cityscapes Dataset, containing urban traffic scenes. The evaluation of the results shows, the necessity to train the FCN-8 on the original size City Scapes Dataset. We conclude that we need to purchase a better hardware.