Visual place recognition using directed acyclic graph association measures and mutual information-based feature selection

Visual localization is a challenging problem, especially over the long run, since places can exhibit significant variation due to dynamic environmental and seasonal changes. To tackle this problem, we propose a visual place recognition method based on directed acyclic graph matching and feature maps extracted from deep convolutional neural networks (DCNN). Furthermore, in order to find the best subset of DCNN feature maps with minimal redundancy, we propose to form probability distributions on image representation features and leverage the Jensen–Shannon divergence to rank features. We evaluate the proposed approach on two challenging public datasets, namely the Bonn and the Freiburg datasets, and compare it to the state-of-the-art methods. For image representations, we evaluated the following DCNN architectures: AlexNet, OverFeat, ResNet18 and ResNet50. Due to the proposed graph structure, we are able to account for any kind of correlations in image sequences, and therefore dub our approach NOSeqSLAM. Algorithms with and without feature selection were evaluated based on precision–recall curves, area under the curve score, best recall at 100% precision score and running time, with NOSeqSLAM outperforming the counterpart approaches. Furthermore, by formulating the mutual information-based feature selection specifically for visual place recognition and by selecting the feature percentile with the best score, all the algorithms, and not just NOSeqSLAM, exhibited enhanced performance with the reduced feature set. •In visual place recognition, nonlinear sequences perform better then linear.•On-the-fly relaxation performs faster than standard shortest path algorithms.•DCNN feature maps represent a place better than hand-crafted features.•Mutual information-based feature selection improves the results in place recognition.