Loop Closure Detection for Visual SLAM Using Simplified Convolution Neural Network

Abstract

Loop closure detection plays a vital role in visual simultaneous localization and mapping (SLAM), since it can reduce the accumulated errors. Handcrafted feature-based methods for loop closure detection have the weakness of lack of robustness with respect to illumination and scale changes. In recent years, the Convolutional Neural Networks (CNN) has been widely used in image recognition due to its feature expressive capacity. However, these deep networks are too complex to satisfy real-time for loop closure detection. In this paper, we propose a novel simplified convolutional neural network (SCNN) for loop closure detection in visual SLAM. We first perform superpixel processing on the original image to reduce the effects of illumination changes. In order to reduce the size of the parameters, we invert the feature maps obtained by convoluting and concatenate them with the original. In addition, we pre-train the proposed network on Places dataset to solve the problem of the scene being unlabeled. Experimental results over CityCenter and NewCollege dataset show that our proposed method can achieve better performance counterparts than other.