A Semantically Aware Place Recognition System for Loop Closure of a Visual SLAM System

Visual place recognition is a core component for relocalization and loop closure detection of a visual SLAM system. Over the last few years, many state-of-the-art methods have been proposed, which mainly use random feature information. However, not all the features equally contribute to the place recognition task, i.e., dynamic objects, textureless area, repetitive patterns. Therefore in this study, we proposed a semantically aware place recognition method that is suitable for efficient place recognition with the meaningful feature. A deep neural network is used to filter out the most meaningful and informative region of an image for better scene understanding. Finally, meaningful features and their geometrical information are extracted and saved in the database for loop detection. For experimental validation, the proposed method has been tested on four well-known public datasets. The extensive experimental results considering all standard evaluation matrices show our method produces superior performance than the other state-of-the-art methods.