DBLD-SLAM: A Deep-Learning Visual SLAM System Based on Deep Binary Local Descriptor

Abstract

The feature quality is critical for positioning accuracy of feature-descriptor-matching-based visual SLAM (Visual Simultaneous Localization and Mapping). State-of-the-art handcrafted descriptors like BRIEF and ORB can hardly work well in complex scenarios. On the other hand, CNN (Convolutional Neural Network) is proved to have tremendous advantages on image feature extraction. In this paper, we develop a CNN model to extract binary visual feature descriptor from image patches with four important loss functions, namely adaptive scale loss, even distribution loss, quantization loss and correlation loss. Based on this learned deep binary feature descriptor sharing the same structure with ORB descriptor, a monocular SLAM system named DBLD-SLAM is designed, by replacing the ORB descriptor in conventional ORB-SLAM. We also train visual Bag of Words to detect loop closure. Experiments show that our DBLD descriptor achieves better results on the HPatches dataset and UBC benchmark. Moreover, the DBLD-SLAM system outperforms other current popular SLAM system on Tartanair dataset.