Detection of Rail-track and Floodwater in UAV Imaging sensors Using Deep Learning

The task of mapping and monitoring water bodies near railway tracks is crucial for railway safety. Accumulation of water near rail-tracks may lead to problems such as washout which may in turn cause accidents and damage to life or goods. Recently, unmanned aerial vehicle (UAV) have gained popularity for monitoring of such water-related hazards around rail-tracks. This research work investigates the effectiveness of a fully convolutional encoder-decoder type network based on U-Net for automated segmentation of rail-track and water regions from UAV-based imaging sensor. Through experimental evaluations using real-world datasets, the performance of the U-Net in segmenting rail-track and water regions is performed. On the Water & Rail-Track (WRT) dataset, the best performance of 0.545 and 0.673 mIoU is achieved for rail-tracks and water classes respectively. The best performance on the challenging Augmented-VOC Dataset is around mIoU of 0.9820 and 0.5283