Automated Semantic Segmentation for Autonomous Railway Vehicles

Abstract

With the development of computer vision methods, the number of areas where autonomous systems are used has also increased. Among these areas is the transportation sector. Autonomous systems in the transportation sector are mostly developed for road vehicles, but highway rules and standards different between countries. In this study, models capable of semantic segmentation have been developed for autonomous railway vehicles with the help of the public dataset. Four different U-Net models were trained with 8500 images for four different scenarios. The model trained for binary semantic segmentation reached mean Intersection over Union (mIoU) value of 89.1%, while the models trained for multi-class semantic segmentation reached 83.2% mIoU, 79.7% mIoU and 29.6% mIoU. Information about the inclusion of high-resolution images in model training and performance metrics in semantic segmentation studies shared.

CM-Unet: A Novel Remote Sensing Image Segmentation Method Based on Improved U-Net

“…Finally, through the full connection layer FC 2 , two feature vectors U avg and U max with the same size 1 × 1 × C are obtained. The specific forms are shown in (1) and (2).…”

“…Potsdam is a typical historical and cultural city with narrow streets and dense residential structures, whereas Vaihingen is a small village with a large number of individual buildings. The GID images cover a geographical area of more than 50000 km 2 .…”

“…Semantic segmentation of remote sensing images plays an essential role in many geographic information applications, such as smart city construction [1], public transportation management [2], and environmental monitoring [3]. Currently, high-resolution remote sensing images are easily accessible, such as unmanned aerial vehicle (UAV) imaging, sensor imaging and radar data.…”