Research On Structure From Motion Method Based On Visual And Inertial Sensors

With the continuous development of technologies such as multimedia communication,machine vision and virtual reality,using multi-view images captured by moving camera to recover three-dimensional object structures is a hot research topic of 3D reconstruction.It can be widely used in the fields of automatic robots,cultural relics protection,medical and medical care,augmented reality and space navigation.The traditional three-dimensional structure recovery method relies too much on geometric calculations,and it does not work well in low-texture,monotonous structures,and so on.In this paper,we study the motion recovery structure method based on multi-vision images and the multi-parameter fusion method based on camera and inertial sensor.In view of the disadvantages of low reconstruction accuracy and lack of scale factor in the present method of motion restoration using monocular camera,this paper proposes and designs a motion recovery structure method based on visual and inertial sensors.The main work and results of this paper are listed as follows:(1)Admitting at 3d reconstruction using monocular camera lacks scale information,this paper introduces the observation data of IMU,and coordinates the camera pose in the time domain and frequency domain.the camera pose is coordinated in the time domain and the frequency domain,and acquired in the frequency domain.The scale information of the monocular camera can solve the problem of the lack of scale of the monocular camera in the 3D reconstruction;(2)The traditional three-dimensional structure recovery method relies too much on geometric calculations,and the effect is not good in the case of low texture and single structure.We apply the convolutional neural network estimating single image depth to the motion recovery structure.The convolution-deconvolution symmetrical network predicts the depth map of a single image,which solves the problem that the output image of the convolutional neural network has low resolution and lacks important feature information.In the experiment,the absolute value error of the depth map reached 0.192,and the accuracy rate was 95.9%,which improved the accuracy of single image depth estimation.(3)Based on the proposed method,this paper builds an experimental platform based on visual-IMU real-scale motion recovery structure method.Finally,the voxel fusion algorithm is used to realize the structural restoration of the sculpture in the corner of the campus,which verifies the effectiveness of the previous method.The reconstructed 3D model has an error of only about 0.2m from the real size,which improves the authenticity of the reconstructed model;it verifies the effectiveness of the previous method.