Intelligent Descriptor of Loop Closure Detection for Visual SLAM Systems

This paper is concerned of the loop closure detection problem, which is one of the most critical parts for visual Simultaneous Localization and Mapping (SLAM) systems. Most of state-of-the-art methods use hand-crafted features and bag-of-visual-words (BoVW) to tackle this problem, but not all SALM systems require hand-crafted feature. With the improvement of machine learning, Convolution Neural Networks (CNNs) has a significant effect on feature detection. This paper proposes a loop closure detection method without hand-crafted feature. We extract the image features through CNNs, and reduce the dimensions of the feature values with t-distributed stochastic neighbor embedding (T-SNE). And then we get a dictionary of two-dimensional feature points, which are obtained by T-SNE. Combined with the new similarity judgment method, the BoVW model based on CNNs is constructed. The new method can solve the loop closure detection of SLAM systems without hand-crafted features. Based on the characteristics of CNNs, the performance of scale-invariant feature transform has been significantly improved.