The Research of Loop Closure Detection Algorithm Based on DeepLabv3+

Abstract

Loop closure detection is the key step of simultaneous localization and mapping (SLAM), which is essential for building a consistent map. Traditional loop closure detection algorithms construct visual bag-of-words by extracting image features to detect whether loops occur. Aiming at the problem that traditional features are easily affected by the external environment and have no ability to understand the environment, a loop closure detection algorithm based on semantic segmentation is proposed. Firstly, perform semantic segmentation on the pre-trained DeepLabv3+ network of the query frame and the candidate frame image, cluster the segmentation results into semantic nodes to construct a semantic topology map, and calculate the similarity of the semantic topology map to measure the spatial similarity of the image. Then, the similarity of the appearance of the image is measured by calculating the similarity of the Hu matrix of each species of static object in the semantic segmentation results. Finally, the species, space and appearance similarity of static objects contained in the integrated image determine the overall similarity of the image for loop closure detection. Experiments show that compared with the traditional bag-of-words model method, the algorithm has more satisfactory performance.