A Robust and Efficient SLAM System in Dynamic Environment Based on Deep Features

In the field of mobile robots, positioning and mapping is one of the most basic problems. A robust and efficient Synchronous Localization and Mapping (SLAM) system is essential for autonomous movement of robots. However, due to the complexity and time-varying nature of the real environment, the positioning and mapping effects will be greatly reduced due to scene changes. At the same time, because of its importance in pattern recognition, deep learning has a relatively mature theoretical foundation and practical framework for feature extraction. In this paper, we propose a visual SLAM system based on deep features in dynamic scenes, which combines mature convolutional neural networks (CNNs) HF-Net into an existing SLAM system. First, use HF-Net to detect the input image, and give local descriptors and global descriptors of the image. Then, these descriptors are used by different modules of the SLAM system. Because the features are not obtained by hand, they are very robust to scene changes. In loop closure detection, a distributed bag-of-words (DBoW) is used to form a vocabulary table, and local and global features are all considered at the same time, so the performance is more reliable. The results show that the entire system has lower trajectory error and higher accuracy on the evaluation data set.