Research Ow Sceve Flow Estimation Method In RGBD Environment

3D Scene Flow is a 3D vector field used to describe the motion of a space object.It is widely used in virtual reality,target detection,target tracking and scene understanding.RGBD scene flow has attracted increasing attention in the computer vision community with the popularity of depth sensor.When performing scene flow estimation in RGBD video streams,the existing segmentation-based methods have better effect on occlusion and large displacement.The layered scene flow method can model the boundary according to the relative depth of known scenes to solve the occlusion problem.The occlusion and scene flow estimation effects of this kind of method will be affected by the segmentation results,and the existing depth sorting process has a low degree of automation.In addition,most of the existing segmentation-based methods perform scene flow estimation under the assumption of rigid motion in the segmentation region,which makes the scene flow estimation of non-rigid body targets inaccurate.In order to solve the above problems,this paper focuses on the 3D scene flow technology,involving the depth image layering technology,scene flow technology with non-rigid motion,and scene flow based on convolutional neural network.The main work and innovations of this article are as follows:(1)A scene flow estimation algorithm based on automatic layering of depth images is proposed.This method firstly uses the depth image repair algorithm to repair the depth image with noise.Secondly uses SLIC superpixel segmentation and similar region merge algorithm to perform the initial segmentation of the depth image.Thirdly adds optical flow constraints to the initial segmentation to achieve the depth image automatic layering.Finally the feasibility and accuracy of the method for calculating the layered scene flow are verified on the dataset Middlebury 2003,SRSF,Princeton Tracking Benchmark,which according to the visual comparison and quantitative analysis of the simulation experiment.(2)A scene flow estimation algorithm based on the local rigid motion assumption is proposed.This method firstly uses the AR(Autoregressive Model)model to calculate the weights of layer support functions in the depth image layering results.Secondly uses a local rigid and global non-rigid assumption method to divide each layer in the depth image layering results into many blocks of the same size.The motion of each block is used to estimate the motion of each layer and the motion of layers is used to estimate the motion of the entire scene.Finally the feasibility and accuracy of the method for calculating the layered scene flow are verified on the dataset Middlebury 2003,SRSF,Princeton Tracking Benchmark,which according to the visual comparison and quantitative analysis of the simulation experiment.(3)A scene flow estimation network model based on convolutional neural network is implemented.This method estimates the scene flow using a encoding-decoding convolutional neural network model.FlyThings3D is used to train the network model for the training dataset.In the training process of the network,the stereo image pair is used as the input of the network,and the scene flow calculated by optical flow and disparity is used as the output of the network.This part saves the time spent using the special network to calculate the depth information;The Monkaa dataset and Driving dataset estimate the scene flow for the test set and verify the feasibility and accuracy of the method.