基于深度学习图像特征的动态环境视觉SLAM方法OA北大核心CSTPCD

Aiming at the problem that traditional visual simultaneous localization and mapping(SLAM)algorithms rely on hand-crafted features,which are not stable enough for dynamic objects and change of illumination conditions and are prone to lose tracking,a stable and real-time method of image feature extraction and matching was presented based on deep learning.The neural network model with attention mechanism was trained through multi-task distillation training to realize feature extraction and matching for scenes with dramatic changes in illumination conditions.Based on the global and local features,a relocalization method based on hierarchical features was proposed to improve the overall accuracy and stability of the system,and it was real-time.Feature extraction and matching tests were performed on images with different illumination and angles in the same scene compared with Superpoint,and localization accuracy tests were performed on TUM datasets compared with ORB SLAM2 and GCN SLAM.Results show that the proposed method can extract sufficient stable features when illumination conditions change dramatically,and it performs better on fr3/sitting_static and fr3/walking_static than the other two methods.The root mean square error of tracks were 6.131 mm and 124.493 mm.Finally,sparse mapping was carried out in real indoor environments,and the effectiveness of the improved relocalization method was verified.