| Computer vision is the simulation for vision formation mechanism of the animal and humanity,and performs the image analysis and understanding in the scene.With the rapid development of computer technology,and the further research on artificial intelligence and pattern recognition,computer vision has been a hot research field.Three-dimensional motion is an important content and hot research in computer vision.It has wide application prospects in high-end video surveillance,vehicle auxiliary driving,3D target detection etc.The scene flow represents the Three-dimensional motion,and the accuracy of scene flow estimation decides the reliability of its application largely,so improving accuracy of scene flow estimation is very important.In this paper,we focus on the topic of scene flow estimation,and apply the variational method to estimate the scene flow.We represent the scene flow as the energy functional and propose a scene flow estimation method based on local rigidity assumption and depth map driven anisotropic smoothing.In the motion estimation,as the local methods are more accurate than the global methods under noise,so in order to improve the accuracy of scene flow estimation,we apply the local rigidity assumption of the scene which means that motion is the same in the local region.However,the local method can only achieve the sparse scene flow estimation,so the total variation(TV)smoothing is also used for the dense scene flow estimation.Furthermore,as boundaries of depth map are aligned with motion edges in most case,in order to constraint the scene flow and improve the accuracy of motion discontinuity preserving,we utilize the boundaries information of depth map to yield TV smoothing.Then the TV smoothing is weighted by depth map driven anisotropic tensor,and we apply the depth map driven anisotropic TV smoothing for the scene flow estimation.After the design of the energy functional for scene flow estimation,an efficient numerical algorithm named the primal-dual algorithm which is a decomposition method is implemented for the variational formulation of the scene flow.The auxiliary variable of scene flow is introduced,then the energy functional optimization can be divided into two parts including the data term based energy functional and the smoothing term based energy functional.For the data term based energy functional optimization can be solved by iterative reweighted least squares with the Gauss-Newton algorithm.The smoothing term based energyfunctional optimization is similar with TV denoising model,it can be solved based on the Legendre-Fenchel Transform.In the experimental section,we qualitatively and quantitatively evaluate proposed method on the Middlebury datasets and compare with other scene flow methods and optical methods.In addition,we evaluate proposed method on real-world scenes acquired by the depth sensor,and observe the estimation results visually.In order to highlight the potential application of scene flow,we also proposed a method of 3D moving object detection.The proposed method clusters and analyses the scene flow of a scene based on the ISODATA,and extracts the moving objects. |
