Based On Direct Method And Visual Dictionary Fusion Slam System Design And Verification

With the rapid development of artificial intelligence,Simultaneous Localization And Mapping(SLAM)has gradually become a research hotspot across computer vision and unmanned driving.In particular,vision sensor-based vision SLAM has become a The main driving force for the development of human driving.Visual SLAM can be divided into feature point method and direct method according to the different estimation methods of inter-frame pose.The feature point method has the advantages of stable feature extraction,strong illumination robustness,and easy detection of closed-loop information.The disadvantages are that it can only build environmentally sparse maps.The direct method is based on the invariant photometric assumption.Through the processing and tracking of pixels,it can be easily constructed The disadvantage of dense maps is that the bag-of-words model cannot be used,which makes it impossible to effectively detect closed-loop information and does not have the ability to actively eliminate accumulated errors?This paper combines the advantages of the feature point method and the direct method,and proposes a SLAM method based on the fusion of the direct method and the visual dictionary,which better solves the trajectory and map points caused by the cumulative error of the direct method when moving in a large outdoor scene.Drift problem.The main research works of this paper are as follows:(1)Summarizes the development of visual SLAM related technologies,introduces the classic framework of visual SLAM,and analyzes the key theoretical technologies involved in visual SLAM.Including motion modeling of rigid body in three-dimensional space,camera imaging model,coordinate system transformation relationship,nonlinear optimization,etc.(2)A visual odometer(D-VO)based on the direct method is designed.The visual odometer consists of a photometric calibration model,front-end tracking,and back-end optimization.The photometric calibration model minimizes the photometric error of the system through photometric calibration and other methods.The front end uses the calibrated photometric information for pixel processing and tracking.The camera pose and map point information are estimated;the backend uses a sliding window and marginalization composed of several key frames to complete the global optimization of camera pose andmap point information.(3)A visual odometer(D-VO)based on the direct method is designed.The algorithm consists of BoVW optimization model,TF-IDF weighting algorithm and closed-loop verification.The BoVW optimization model uses the M-ORB feature point extraction algorithm and a hierarchical dictionary tree structure,which greatly reduces the calculation amount and improves the real-time performance compared with the original model;the TF-IDF algorithm provides reasonable weights for each visual word;closed-loop verification Guarantee the accuracy of the algorithm.Finally,the superiority of the algorithm is verified on the TUM public data set,especially in the face of large-scale outdoor complex scenes,without reducing the accuracy and robustness,it still has high processing speed and low operation.time..(4)A SLAM system(D-SLAM)based on direct method and visual dictionary fusion was proposed,and the construction of a visual SLAM multifunctional experimental platform and corresponding experimental verification work were completed.First,the improved closed-loop detection algorithm was integrated with the direct method visual odometer to obtain the direct method SLAM system(D-SLAM).Then,based on the Ubuntu system,the visual SLAM multifunctional experimental platform required for this paper is designed and built.Finally,on this platform,D-VO and D-SLAM systems and other mainstream SLAM systems were tested under different scenarios and closed-loop experiments,and a quantitative analysis of the error function was given.The experimental results show that the D-SLAM proposed in this paper has the ability to accurately build a map of the surrounding environment,especially in the face of large-scale outdoor complex scenarios,it can actively eliminate the accumulation of errors and accurately build a globally consistent map..