Improved Deep Distance Learning for Visual Loop Closure Detection in Smart City

Abstract

Visual Simultaneous Localization and Mapping (vSLAM) are expected to promote the initiatives in Smart City including driverless cars and intelligent robots. Loop closure detection (LCD) is an important module in a vSLAM system. Existing works with convolutional neural networks exhibit better performance on feature extraction, but this is far from enough. Concerning the characteristics of LCD, it is of great significance to have a customized loss function and a method to construct suitable training image sets. Based on this motivation, we propose a novel framework for LCD. Through a deep analysis of the distance relationships in the LCD problem, we propose the multi-tuplet clusters loss function together with mini-batch construction scheme. The proposed framework can map images to a low dimensional space and extract more discriminative image features, which help learn a more essential distance relationship of the LCD problem. Extensive evaluations demonstrate that our method outperforms many state-of-art approaches even in complex environments with strong appearance changes. Importantly, though the training process is computationally demanding, its online application is very efficient.