基于深度学习的SLAM算法研究

Vision-based simultaneous localization and mapping (VSLAM) is currently an important research direction in the fields of 3D reconstruction and computer vision, and is the core technology in the fields of autonomous driving androbotics.In recent years, deep learning has advanced by leaps and bounds, demonstrating powerful capabilities in areas such as object tracking, object detection, and image recognition. In order to explore the performance of deep learning in the field of VSLAM, this paper uses feature points based on deep learning and traditional descriptors that draw on deep learning training techniques to improve the accuracy and robustness of SLAM.This paper takes ORB-SLAM2, a classical framework in the field of visual SLAM, as the baseline, uses HFNet feature and Box Average Difference (BAD) descriptor to enhance the performance of the SLAM. The results tested on publicly available datasets show that the improved system can significantly improve the accuracy and robustness of the system while ensuring real-time performance, demonstrating the superiority of deep learning-based networks and their advanced techniques over traditional methods.