Research And Implementation Of Deep Learning Based Visual SLAM Technology

Robot vision SLAM technology has a wide range of applications in autonomous driving,navigation and positioning.However,in some special situations(such as strong lighting),robot vision SLAM often has problems such as scale drift and scale error due to insufficient accuracy.The reason is that most of the existing visual SLAM algorithms need to manually design sparse image features,while artificially designed sparse image features often contain certain assumptions about the environment(such as sufficient illumination,material,etc.).This causes the environmental depth information estimation effect to deteriorate when environmental factors such as lighting conditions change.Deep learning has replaced the artificial design features and feature extraction methods with unsupervised or semi-supervised feature learning,and has achieved good results in many fields such as image processing.In this paper,the robot vision SLAM is taken as the research object,and the deep learning method is taken as the basic method to study the accuracy of SLAM in special scenes.In this paper,the monocular image depth estimation method based on camera pose conversion relationship is proposed.The deep learning method is used to estimate the depth information of the environment in the image.Then the unsupervised learning depth estimation aided ORB-SLAM algorithm is proposed.The deep learning method proposed in the previous step is integrated into the existing ORB-SLAM algorithm.The main work of this paper is as follows:(1)A monocular image depth estimation method based on camera pose conversion relationship is designed.This method draws on the camera pose estimation method in the traditional SLAM system to enhance the supervision signal and increase the training constraints.In the camera pose estimation stage,a pose image is constructed by establishing a continuous image window,and the pose transformation matrix is calculated by using the pose transition relationship,and the pose map is improved.The pose map can partially eliminate the cumulative error,so the accuracy of the image depth estimation can be improved compared to the existing algorithm.(2)The unsupervised learning depth estimation aided ORB-SLAM algorithm is designed.This method combines the traditional ORB-SLAM algorithm with the monocular image depth estimation method based on the camera pose conversion relationship proposed in this paper.When the traditional ORB-SLAM cannot match enough feature points,the deep learning method is used to estimate the environmental depth information.This method can complement each other in the environment of weak texture and strong illumination,and can effectively improve the accuracy of the environmental depth information,and indirectly improve the accuracy of the ORB-SLAM algorithm for modeling the environment.Based on the above implementation schemes and mechanisms,this thesis designs and implements a deep learning-assisted visual SLAM prototype system,and conducts experiments in public datasets and real-world scenarios,and comprehensively verifies the work.