Research On Loop Closure Detection Of Visual Slam Based On Extraction And Matching Of Local Features

SLAM(Simultaneous Localization and Mapping) algorithms perform mapping and localization tasks at the same time, creating an incremental map of an unknown environment while localizing the robot in this map. These techniques are essential for achieving autonomy in mobile robots. In SLAM, loop closure detection is a key challenge to overcome which entails the correct identification of previously seen places from sensor data, allowing the generation of consistent maps. Due to the low cost of cameras, the richness of the sensor data provided and the availability of cheap powerful computers, loop closure detection of visual SLAM have recently received wide attentions.However, the shortcomings and inevitable calculating errors of modeling in visual information acquiring, describing and matching cause great difficulty in the mobile robot to extract the precise loop closures and thus hinder the SLAM tasks to be completed successfully. The visual loop closure detection is still one of the most challenging problems in outdoor environments. In this paper, we study the SLAM based on monocular vision sensor and detect loop closures by matching a subset of previously collected frames when using color images. The principle and various factors that govern the performance of Bag-of-Visual-Words(Bo VW)method are analyzed in loop closure detection. The major work of this paper is expressed in the following :First of all, we analyze the feature detection and compare KAZEfeatures with the state-of-the-art algorithm, SIFT. Bo VW scheme are formed through the extraction of the features in graph-based SLAM. We analyze how to tailor visual vocabulary generation so as to yield more discriminative visual words. The discriminative visual words depend on robust feature detection which improve matching performance of loop closure detection.Secondly, there are kinds of methods of image matching described in this paper and the comparative analysis of KAZE, SIFT and KAZE+SIFT in the different date set. It is concluded that KAZE possesses strong robustness to brightness, scale, noise. So we chooses KAZE for operator feature extraction in the next framework.Finally, KAZE is optimized based on the above analysis, and a robust framework for the loop closure detection based on KAZE is proposed. We achieve the dependencies by learning a tree-structured Bayesian network using the Chow-Liu tree, which yields the optimal approximation to the joint distribution over word occurrence within the space of tree-structured networks. Then, we take advantage of the scalability of inverted index techniques. It can quickly search the scene containing the specified words and avoid comparing with the previous frames. The loop closure detection algorithm is applicable in the dictionary with large capacity. The framework permits efficient learning and inference even for very large visual vocabulary sizes. Outdoor scene experiment results show the effectiveness and the feasibility of this algorithm.