Research On Robot Robust Vision Localization Method For Low-Texture Scenes

Robots are widely used in diverse scenarios such as home and industry.With the development of technology,cameras have gradually become the core sensors of home service robots due to their advantages of low cost and rich information.However,in the real environment,low-texture scenes such as corridors,large white walls,and ceilings cause problems such as difficulty in visual feature tracking and errors in loop closure detection for the robot’s visual positioning and navigation.Aiming at these problems,this thesis conducts research on visual feature tracking for low-texture scenes to improve the robustness of visual odometry;carry out research on loop closure detection based on sequence verification to achieve accurate low-texture loop detection;and construct lowtexture scene visual positioning System,to verify the robustness and practicability of the system.The main research content of this thesis is as follows:(1)Aiming at the low robustness of feature tracking and the small number of tracking feature points in the visual localization system in low-texture scenes,this thesis proposes an indirect feature matching method based on auxiliary information.First,a large number of basic matches are established between frames,and then the correct match is screened out through the adjacent support matching strategy,which solves the problem of sparse matching caused by direct matching,improves the number and quality of feature point matching,and achieves better results in low-texture image matching.good result.At the same time,in order to solve the problem of false matching caused by repeated textures in low-texture scenes,this thesis intends to introduce motion information to constrain the feature matching area,guide the direction of feature search,and avoid false matching caused by repeated textures.By solving the problem of stable feature tracking and repeated texture mismatch in low-texture scenes,the feature tracking method in this thesis improves the positioning accuracy and robustness of the positioning system in lowtexture scenes,and can run in real time.(2)Aiming at the problem that images of different locations in low-texture scenes are judged as loop closures,this thesis proposes a loop closure detection method based on sequence matching.In the process of similarity score calculation,this thesis introduces similarity score suppression methods within images and between images.Reduce the weight of high-similarity features in the similarity score,and use Bayesian inference to perform loop closure judgments on the similarity score with the help of previous and subsequent frames.In addition,in order to deal with the extreme similarity of artificial scenes,based on the assumption of similarity in image sequences observed during motion,the false positives caused by the high similarity between frames are further filtered out according to the length of similar sequences.In low-texture scenes,the method in this thesis improves the recall rate by an average of 22% when the accuracy rate is 100%,and shows higher precision and robustness in low-texture scenes.(3)In order to verify the effectiveness and practicability of the proposed robust positioning method for low-texture scenes,this thesis builds a low-texture scene visual positioning system based on monocular vision and IMU,and uses an unmanned vehicle platform for data collection.Robust positioning of the robot is completed under the following conditions,and the situation of system crashes has been significantly improved.The average error of APE on the low-texture dataset is 0.220678 m,which is only 14.6%of that before improvement.At the same time,it can also filter out false positives for loop closure detection in low-texture scenes.Experimental results demonstrate the superiority of our method for localization in low-texture scenes.