Loop Closure Detection for Visual SLAM Based on SuperPoint Network

摘要

Loop closure detection is one of the important components of Visual SLAM. It can reduce the accumulating drift of visual odometer if the loops are detected correctly. Traditional loop closure detection mainly relies on the traditional algorithms to extract feature points. However, these algorithms are difficult to adapt to the complex and dynamic environment, e.g. illumination changes. In this paper, a new approach of loop closure detection is proposed. It learns inner structures from raw data using neural networks named SuperPoint. SuperPoint is designed to detect interest points and corresponding descriptors at the same time. The method clusters the feature descriptors as words using K-means and converts descriptors of each tested image into a vector by comparing the similarity between the descriptors and words. It measures the similarity of the images by calculating the cosine similarity of the corresponding vectors. A series of experiments on New College and City Center are demonstrated to evaluate the performance of the proposed method. The average precision of loop closure detection based on SuperPoint increases by 28% and 122% respectively compared with that based on SIFT and ORB under the same conditions. It is proved that the method based on SuperPoint is possible to detecti loops at a satisfactory precision compared with the traditional method based on ORB or SIFT.