Relative Attitude Perception Method Of Capsule Robot Based On Computer Vision

In the current field of intestinal disease inspection,traditional endoscopy can no longer meet people's needs due to its high pain and high risk characteristics.The use of non-invasive capsule robots to enter the human body for inspection is becoming a new type of inspection method.In the actual diagnosis and treatment process,in order to facilitate the doctor to observe the lesion,the feedback control information of the capsule robot's attitude is very important.The existing attitude sensing methods mostly use signal detection or sensor array measurement,which greatly consumes hardware resources and increases costs,which is not conducive to the popularization and use of capsule robots.This thesis starts with the camera carried by the capsule robot,and proposes a method of capsule robot attitude perception based on computer vision.The method is divided into four modules: feature enhancement module,fold matching module,optical flow point tracking module and posture calculation module.The visual invariant features in the intestine can be switched between fold matching and optical flow point tracking according to the strength of the intestinal villi information.This thesis first established multiple coordinate systems in the magnetic field control system,and deduced the pose description method of the capsule robot in the world coordinate system.Then,on the basis of fully analyzing the anatomical characteristics of the physiological structure of the intestine,intestinal folds and villi were selected as visual invariant features.Considering the lack of light in the intestines,this thesis uses single-scale Retinex to compensate for the illumination of the image.At the same time,bilateral filtering is used to reduce noise while retaining as much as possible the fold information in the intestines for subsequent extraction and matching.For semi-dense intestinal folds,this thesis uses adaptive threshold segmentation to extract folds.On this basis,the matching of folds is divided into two parts: coarse matching and precise matching.Coarse matching mainly solves the overall matching problem of intestinal folds.Therefore,we start with the Hu moment with RST invariance,and use the Euclidean distance of Hu moments between different intestinal folds as the similarity measure.The precise matching process is to obtain the corresponding relationship between the intestinal fold points,so we use the shape context to abstract it into a weighted bipartite graph matching,and use the Kuhn-Munkres algorithm to solve it.When the feature points in the intestinal picture are larger than the threshold,this thesis adopts the Lucas-Kanade optical flow method with stronger real-time performance.The motion information of the capsule robot is decomposed into the optical flow increment of each layer on the image pyramid,and the solution is iteratively solved from top to bottom.The traditional Shi-Tomasi corner point extraction algorithm is improved,and the extracted corner points are used as the optical flow points in the initial frame of the intestine.After obtaining the matching relationship between the fold point or the optical flow point in the front and rear frames,this thesis divides the capsule robot's posture calculation into posture calculation and optional posture optimization.First,the motion parameters of the capsule robot's posture change are solved through the epipolar geometric constraints,and the relative depth information of the spatial points in the intestine is recovered based on this.Then the intestinal space points and motion parameters are used as variables to be optimized,and the bundle adjustment method is used to optimize.In order to verify the feasibility of the posture perception method,this thesis conducted experiments in an ideal intestine model,and measured the accuracy of posture calculation by the reprojection error of the spatial point in the intestine.Experimental results show that the posture calculation error of this sensing method is less than 5 pixels,which can meet the needs of the capsule robot for posture perception in the human intestine.