PanoNetVLAD: Visual Loop Closure Detection in Continuous Space Represented With Panoramic View Using Multiple

Abstract

This study deals with the problem of visual loop closure detection that occurs when there are limited camera field of views. The proposed method generates a panoramic image using multiple images and creates sub-panorama images through the sliding window process. Then, a learning-based descriptor is formed for robust feature detection even in illumination changes and challenging environments. After that, the index of the sub-panorama image descriptor, which is closest to the query image descriptor, is found, and the final candidate is found which has the same index with the image of the raw database. Histogram equalization and image masking were conducted to create a robust panorama image before examining our proposed method. To prove the effectiveness of our proposed algorithm, a dataset trajectory was acquired in which visual loop detection was difficult. In addition, datasets were obtained indoors and outdoors to check whether the proposed algorithm is applicable to various environments. The acquired dataset is shared through the GitHub page at https://github.com/SungJaeShin/KAIST-DP.