Qualitative motion analysis using a spatiotemporal approach.

Abstract

In computers, visual motion perception is a process which transforms a sequence of digitized two-dimensional images into a three-dimensional description of a scene in terms of objects, their three-dimensional shape, and their motion through space. Conventional approaches require the image flow to be extracted first. However, obtaining exact image flow vectors has been considered a complex task by most dynamic vision researchers. In this dissertation, we show that, using spatio-temporal approaches, qualitative motion information of various kinds can be obtained from image sequences easily and reliably. A spatio-temporal 3-surface model is first presented and, based on this model, we analyze mathematically how the geometry of the intensity hypersurface gives information about motion in images and we have shown that the validity of this constraint equation has nothing to do with the constant brightness assumption. The second part of this thesis continues the use of this 3-surface model. We apply the model to the three-dimensional image segmentation problem. As opposed to conventional approaches which segment each 2D image based on motion analysis results, our approach performs motion analysis based on the 3D segmentation result. We show that any spatio-temporal image solid can be segmented into a set of disjoint volumes each of which is a trace of two-dimensional regions. The entire grouping process in the spatio-temporal space does not require any motion information to be pre-computed. We present a parallel algorithm for the general 3D image segmentation problem. Finally, we show that certain qualitative motion information can be obtained without the computation of the image flow field. Each segmented volume found by the 3D segmentation algorithm can be identified as corresponding to the area of either a moving or a stationary object in the scene. For any volume identified as corresponding to a moving object, its direction of motion in three-dimensional space with respect to the camera is further classified into one of the ten ranges of angles. Experimental results demonstrate the effectiveness of our algorithms on several image sequences including both outdoor and indoor scenes.