Motion Segmentation And Dense 3-D Interpretation Based On Monocular Image Sequences

3D moving object detection and reconstruction is an important research topic of computer vision. This paper will study how to achieve a unified framework for 3D monocular image sequences of moving objects segmentation, motion estimation and depth estimation.In this paper, a new method is proposed for motion segmentation and the dense 3-D interpretation based on monocular image sequences. The method achieves spatiotemporal process model, using Dominant Motion Segmentation method, Active Contour Model, Variational Method and 3-D motion Model. This model utilizes modern mathematical methods to study the 3-D motion segmentation and the dense 3-D interpretation as a whole and convert it into a unified segmentation optimization and estimation problem.First, an energy function including motion information, depth information, and segmentation surface is proposed based on the maximum a posteriori, motion and depth constraint model, observation model and prior model.Secondly, in this paper the energy function is minimized through Variational method and gradient descent method. The motion and depth are estimated through Numerical Partial Differential Equations.Thirdly, the depth information is corrected through the gray image and the corresponding color image or gray image. And more accurate depth information is achieved. The experimental results show that the algorithm can simultaneously achieve the motion segmentation of monocular image sequences and the dense 3-D interpretation, and can handle the case of moving multi-regions and movements of both viewing system and objects.At last, we give the summary of the algorithms in this paper and the future research directions.