수중 소나 영상의 부분 왜곡에 따른 학습 데이터 Augmentation을 통한 마커 검출 성능에 관한 연구

Abstract

The existing SLAM study refines the position of the mobile robot by using the landmarks obtained from the environment based on GPS. However, since it is impossible to use GPS in an underwater environment, detection of landmarks by sensors becomes very important. Unfortunately, the use of artificial landmarks is necessary because few features make it a natural landmark in a typical aquatic environment. The purpose of this study is to detect artificial markers based on thedeep learning technique robustly. It always does not guarantee good results to use a complex deep-learning model, so it is needed to find the best model by adjusting the layers to get the best performance. In addition, the recognition rate of the deep-learning model is reduced by several noise such as distortion etc. during data acquisition. To solve this problem, the training data augmentation for the distortion was executed with the rotation. In this paper, we apply the object detection for the sonar image data of three types of artificial markers by using the Faster R-CNN.