Research On Vision-based Positioning And Composition Algorithms

SLAM(Simultaneously Localization And Mapping)is intended to restore the position of the robot and create a 3D map for the environment while moving through the sensor it carries.This problem has been researched for more than 30 years since it was proposed in 1986.At the same time,the algorithms of Localization and Mapping have evolved from the framework based on laser sensors to the visual sensor.The laser-based SLAM is proposed earlier,the theoretical method and the computational framework of the laser-based SLAM have been formed.Currently,the research on SLAM is mostly focused on the vision-based SLAM.In the vision-based SLAM,the direct image alignment model is usually used to calculating the translation and rotation between two images.The principle of this model is to assume that the light value of a point in the 3D space remains the same value in a short time and at a small angle.The relationship of the brightness error between projection points is established by this constraint relationship.And the translation and rotation between cameras are solved by the minimizing brightness error.In the experiment,we find that the traditional direct method has a large error while the camera rotates quickly.That is,when the separation distance between image frames increases,the matching error increases significantly,and even serious mismatching occurs,which is the limitation of the traditional direct method.In this article,the reason why the traditional direct method model fails to align images when camera moves fast is that there is no reliable iterative initial value in the nonlinear optimization process.It can be seen from the theory of nonlinear optimization,the nonlinear optimization depends heavily on the iterative initial value.And improper selection of the iterative initial point will result in the algorithm falling into a local minimum.However,the direct method model does not offer an effective way to calculate the initial value for the iterative algorithm,which makes the optimization fall into the local minimum.Based on these,we propose a Novelty model based on priori estimation of the inertial measure units(IMU):(1)Firstly,we use the inertial measurements to calculate the rotation and translation matrix that are relative to the reference of camera position at the previous moment.(2)Secondly,we use this value as the initial value of nonlinear optimization in the process of direct image alignment and obtain accurate camera pose by minimizing the photometric error.(3)Thirdly,we build Bayesian filter to correct the bias of IMU sensors.In the chapter four,We extend the this model into full SLAM system,which is real-time dense point cloud map building system based RGB-D camera.Finally,we evaluated the LSD-SLAM framework,the ORB-SLAM framework and our own framework on TUM datasets and ourselves datasets recorded in lab environment.The results show that our system can runs at twice the speed of the ORB-SLAM system while maintaining the same positioning accuracy as the orb-slam framework.Moreover,compared to the other frames,our system has the function of building a dense point cloud map.