Fast and Robust Visual Loop Closure Detection with Convolutional Neural Network

Abstract

Visual loop closure detection has always been a major problem of visual SLAM. When faced with various scenarios, the accuracy of the creation and positioning of the map by SLAM depends on the performance of the loop closure detection algorithm. It is precisely because the SLAM system is affected by the cumulative error when creating the map, so the loop closure detection algorithm research has become the current hot research direction. After years of research, it is found that the traditional closed-loop detection algorithm does not perform well when the time, light, and weather changes encountered in the scene recognition process. And the speed is slow, it is difficult to run in real time on the mobile terminal (such as mobile robot). Now, due to the research of deep learning, it is possible to improve the visual loop closure detection algorithm. First, this paper uses the improved pre-training model Lite-shuffleNet network to extract the depth and semantic information of the image to obtain the feature descriptor; then calculate the cosine similarity; finally a pair of best candidate frames are obtained, judged as a loop. The experiment uses the City Center dataset and the New collage dataset to verify the algorithm. The experiment shows that the algorithm proposed in this paper is superior to the traditional loop closure detection algorithm and the VGG16 deep learning algorithm in performance, and has a better Precision-Recall rate and feature extraction speed.