Design And Implementation Of Motion Estimation Software Based On TOF

Scene flow estimation is a hot area in both scientific and industrial fields,which plays an important role in 3D scene understanding and scene dynamic reconstruction.Scene flow describes the three-dimensional motion field formed by relative motion between the scene point and the camera.Its essence is the expansion of optical flow in three-dimensional space.Compared with optical flow,scene flow can more realistically and comprehensively represent the motion field of the scene and provide more dynamic scene information for the computer vision system.However,due to the large spatial scale of the solution,the computational cost of scene flow represented by dense motion field is huge,so there are few methods for real-time solution.In addition,the scene flow is often visualized by the corresponding optical flow image,but this way cannot visually display the three-dimensional motion field of the scene.Therefore,this paper designs and implements a motion estimation software based on time-of-flight camera,which is divided into image reading module,scene flow estimation module and visualization module.The image reading module is mainly used to read RGB-D video frames collected by TOF camera.For the scene flow estimation module,this thesis proposes an improved scene flow estimation algorithm based on RGB-D image sequence,which can jointly estimate the camera motion and scene flow.The key of the algorithm is the two-fold segmentation of the scene.First,the scene segmentation algorithm based on K-Means is used to divide the scene into geometric clusters and treat each cluster as a rigid body,which greatly simplifies the estimation of scene flow.Then,the scene is divided into the background region and the moving region.All the clusters divided into the background are used to refine the camera motion,and each moving cluster calculates independently its rigid motion to obtain the corresponding scene flow.Comparing the proposed algorithm with several state-of-the-art traditional methods in scene flow estimation,the results show that the method can improve the accuracy of scene flow estimation while maintaining high real-time performance.For the visualization module,this thesis creates a VTK dataset based on the scene point coordinates and the corresponding scene flow and builds a VTK visualization pipeline to realize the visualization of scene flow.This can help users make accurate judgments on the motion state of moving objects,and also provides a basis for the optimization of scene flow estimation algorithms.