An improved deep-learning monocular visual SLAM method based on local features

Abstract

To improve the tracking performance of the Simultaneous Localization and Mapping (SLAM) system, this paper presents a monocular visual SLAM method based on deep learning second order similarity of local features. The local features are generated into descriptors from patches around key points in a frame using a deep neural network. It is applied to tracking, re-localization, and loop closure module to enhance data association. We also train a visual bag of words model to adapt to the local descriptors. Additionally, we use two adaptive strategies to improve the proposed method, one strategy refines key points detection with illumination intensity, and the other strategy reduces the possibility of tracking lost based on the ratio of outliers’ number in feature matching. We evaluate our method on two public datasets. The experimental results demonstrate the effectiveness of the system and also show that the adaptive strategies can increase tracking performance and improve the robustness in challenging conditions.