Semantic Segmentation of River Video for Efficient River Surveillance System

Abstract

Development of an efficient river monitoring system with state-of-the-art AI technology becomes more and more important. Nevertheless, training a network requires a large amount of annotated datasets which is an exhausted work. In this work, an original dataset is generated by modular interactive video object segmentation(MiVOS) including daytime and nighttime surveillance video of Inoo River in Tokushima. Then, an efficient segmentation tool named Hyperseg is trained by the dataset. The semantic segmentation results of the river video show that over 0.8 of IoU is obtained on the classes of bridges, rivers, and banks. It is considered an acceptable segmentation performance to estimate the water level of the river to construct a smart river surveillance system.