D_ORB: A Robust Visual SLAM System Based on Deep Feature Extraction

For the visual Simultaneous Localization and Mapping (SLAM) algorithm, although a good framework has been established in most aspects, feature extraction and association are still empirically designed and may be fragile in complex environments. The integration of deep convolutional neural networks (CNNs) for feature extraction into the SLAM framework is proposed in this article. The system utilizes state-of-the-art CNNs to detect keypoints in each image frame, which are then evenly distributed through octree filtering. At the same time, CNNs generate corresponding descriptors for these keypoints. These CNN-derived features are integrated into various SLAM modules, demonstrating superior robustness to environmental variations and viewpoint changes compared to traditionally crafted features. Additionally, the bag-of-words (BoW) method is employed to train visual vocabulary features, and based on these features and vocabulary patterns, a reliable loop detection mechanism is established. The experimental results show that all proposed modules are significantly better than the benchmark model, and the entire system achieves lower trajectory errors and higher tracking robustness on all evaluation datasets.