Training a robust UAV river patrol AI for different river and an analysis of the training dataset

Abstract

Japan has a lot of river. For the efficient river patrol, it is planned to use UAV & AI. The river patrol must detect the various target from the various back ground. This study focuses on the aerial image on the two river in urban or mountainous area. The target such as illegal dumping or people, bicycle, and cars can have the different distribution, and it is considered that the distribution of them and the variety of back grounds can affect the AI detection precision. The training uses the Faster R-CNN. The dataset is composed by the data on each river only or mixed with the ration changed. In this study, by the precision of the models are verified, the possibility of training models robustly for the different river and the dataset are verified.