Optical Snow Analysis using the 3D-Xray Transform

There are many methods to analyze motion in computer vision. Most of the classical methods use optical flow, layered motion or segmentation. Optical snow is a complex motion estimation scenario which analyzes motions such as snowfall, tree movements, cars traffic and people walking. These scenes are made of many different objects, moving in different speeds in various directions. While analyzing these scenes, we have to recover both the motion and object segmentation. In this paper, we detect optical snow motion in video by using the discrete 3D-Xray transform. The algorithm uses the discrete 3D-Xray transform, which is situated as a core algorithm in medical imaging when 3D reconstructions from projections are needed, to partition the captured data and to assemble 3D space into specific planes. The detection of optical snow motion with object tracking are achieved through analysis of the energy distributions on these 3D-Xray planes. The output from this analysis contains local and global directions of the movements of these objects. The algorithm identifies and preserves the path of tracked objects even when there are multiple objects, or they are partially covered (occluded) by other objects or there is a compound object of merge and split as players in a soccer game or as in surveillance applications. The algorithm analyzes frame by frame a video sequence without the need to have neither object segmentation nor clustering. The algorithm is exact and geometrically faithful as it uses summation along straight geometric lines without any interpolation schemes. In addition, it detects the direction of each object and computes its relative velocity. The algorithm has two limitations; the moving objects have to move in straight lines with constant velocity. The algorithm is computationally efficient since the computation utilizes only certain planes without the need to have an exhaustive search and computations of of the activities in all planes.