Loop closure detection for visual SLAM systems using deep neural networks

The detection of loop closure is of essential importance in visual simultaneous localization and mapping systems. It can reduce the accumulating drift of localization algorithms if the loops are checked correctly. Traditional loop closure detection approaches take advantage of Bag-of-Words model, which clusters the feature descriptors as words and measures the similarity between the observations in the word space. However, the features are usually designed artificially and may not be suitable for data from new-coming sensors. In this paper a novel loop closure detection approach is proposed that learns features from raw data using deep neural networks instead of common visual features. We discuss the details of the method of training neural networks. Experiments on an open dataset are also demonstrated to evaluate the performance of the proposed method. It can be seen that the neural network is feasible to solve this problem.