Proposing a Novel Multi-Frame Motion Estimation at Occlusion Conditions

The field of video motion estimation plays an important role in computer vision and is one of the dominant issues in this realm. Optical flow is one of the most famous approaches to this problem. Changes in the 2D projected images are caused by the relative movement of a 3D scene using a camera. The computation of a projection of the actual motion onto the picture plane is used to estimate this motion. The optical flow is a 2D displacement field that describes the apparent motion of brightness patterns between two successive images. Several solutions have been proposed to solve the optical flow problem. Occlusion handling constitutes a significant problem in the field of optical flow estimation. Image warping and feature warping plays an essential role in optical flow; however, occlusion is a major source of error. We propose a framework as a learnable occlusion handling in motion estimation. We focus on the feature matching process in warping which is a core technique in optical flow estimation and we demonstrate that it helps filtering the occluded pixels while estimating the optical flow. The occlusion module can be integrated into other end-to-end networks in different computer vision tasks including object detection and segmentation. The results show that the proposed occlusion-sensitive frame for optical flow outperforms existing methods.