Research on the Improvement of ORB-SLAM2 Algorithm Based on Deep Learning

Vision-based Simultaneous Localization and Map Building (VSLAM) refers to the camera as the only external sensor that creates a map of the environment while positioning itself. It is an important research direction in the field of 3D reconstruction and computer vision, and a core technology in the fields of autonomous driving and augmented reality (AR). At the same time, with the popularity of mobile devices and the improvement of computing power, mobile augmented reality has shown high practical value. Visual SLAM has developed rapidly over the past decade. The system can run on micropcs and embedded devices, and even on mobile devices such as smartphones. Aiming at the shortcomings of SLAM, this paper improves the algorithm and engineering aspects of the core module of SLAM system, proposes an efficient and robust visual SLAM system, and implements the augmented reality application based on SLAM algorithm on the mobile terminal platform. This paper mainly does the following work: summarize the development status of SLAM, analyze the theoretical application and future development trend of the combination of SLAM and deep learning, use deep learning-based feature points to improve the accuracy and robustness of SLAM, take the classic framework of visual SLAM 0RB-SLAM2 as the baseline, and improve ORB-SLAM2 system implements and uses HFNet feature points to improve the performance of SLAM systems.