Research Of Visual Gyroscope-Aided Pedestrian Dead Reckoning

Pedestrian Dead Reckoning(PDR)is an indoor positioning method of calculating relative position by observing changes in heading and distance.It is usually implemented with inertial measurement unit such as accelerometers,magnetometers,and gyroscopes that do not require external facilities.However,due to the inherent properties of the inertial sensors,PDR may have an accumulated error that is difficult to estimate under longterm and long-distance conditions,resulting in limitations on the use of PDR.Therefore,how to eliminate effectively and stably the cumulative error of PDR has become a research hotspot.In recent years,more and more auxiliary facilities or means such as WIFI,Bluetooth,etc.have been used to eliminate the cumulative error of PDR.However,these aids increase the dependence of the system on external facilities to a certain extent.With the rapid development of Micro-Electro-Mechanical System(MEMS),mobile devices such as smart phones and tablets that people carry with them are mostly integrated with micro-inertial measurement units and high-resolution cameras.In recent years,computer vision-based navigation and positioning technology has developed rapidly,and the most representative of which is Simultaneous Localization and Mapping(SLAM).The camera and inertial sensors have natural complementarity from the perspective of navigation and positioning.For example,camera is greatly affected by changes in the external environment.In the process of visual SLAM,the frame rate of the image is low and the accumulated error is relatively small.And when the spatial feature is stable,the estimation of pose is relative accurate;the inertial sensors have a high operating frequency and adopts high frequency the integration of measurement is more accurate in a short time.And beside this,it is less affected by the outside world,but the long-term error accumulation is relatively large.In this paper,a visual aid-based pedestrian voyage estimation method(Vision-aid PDR,VPDR)is proposed considering the complementary characteristics of the two sensors.The accumulated error of the PDR is reduced or eliminated by the fusion of the camera of the smartphone and the inertial sensor in an indoor scene.Among them,unlike the traditional visual positioning method to detect and track feature points,this paper attempts to detect the structural lines in the scene and calculate the vanishing point to track the camera's attitude.There are two advantages to this approach: firstly,there are weak textures or textured-less environments(such as corridors,walls,etc.)in indoor scenes,which make traditional visual positioning methods unable to detect feature points and thus cannot work stably and robustly;Secondly,the proposed recovery of the camera pose from structural lines is less computationally complex than the traditional feature-based approach.The process of recovering the camera from the vanishing point is roughly similar to the principle of recovering angles from the gyroscope,so we abstractly call it a "visual gyroscope."The main research contents of this paper include:1.Design of visual gyroscope based on vanishing pointDue to the principle of projective geometry,the lines parallel to each other in the camera coordinate along to the propagation will be intersected to a point which called vanishing point.The change of the orientation of the camera can be derived according to the change of the vanishing point on the image plane.This is the theoretical basis of the visual gyroscope.The stability and accuracy of vanishing point directly affect the performance of visual gyroscope.Therefore,this paper designs a two-stage vanishing point filtering algorithm: the first stage we utilize statistical methods for the filtering of all candidate vanishing points,and the second stage we design a Kalman filter combining inertial measurement units and candidates vanishing points to fuse to obtain a more accurate vanishing point.2.Research on fusion algorithm of PDR and visionThe change in heading angle in the image coordinate system obtained by the visual gyro.The PDR obtains the heading of the descendant of the sensor coordinate system.In this paper,the method of Kalman filtering is used to fuse the heading angles of the image coordinate system and the inertial coordinate system to obtain a more accurate heading angle.3.Position recognition and pose estimation based on line descriptorSimilar to feature points,the line features of an image also have corresponding descriptors.This paper presents an extended binary line descriptor ELBDP.In the VPDR process,the candidate images with absolute poses in the database are matched in real time.After initialization,we proposed a lightweight line-matching method to recover the relative pose of the image and use it in VPDR to further reduce the cumulative error of VPDR.