Online motion segmentation through multi-temporal section motion analysis in dynamic scenes

Abstract

In this dissertation, we deal with the problems of online motion segmentation in dynamic scenes and its application to generic object hypotheses generation for window-based object detection.

First, we present an algorithm of online motion segmentation that runs rapidly with low computational complexity for processing streaming images. For rapid operation with low computational complexity, we represent feature points tracked over the frames, using a spatial neighbor graph. Furthermore, for multiple-frame observation of the feature points, we utilize a sliding temporal window, and compute the affinities for each point pair on the graph, through motion analysis based on multi-temporal sections in the sliding temporal window. Then, these points are optimally segmented, producing motion-segmented points.

Second, we tackle the problem of online motion segmentation in highly dynamic 3D scenes. In order to generate trajectories of feature points in highly dynamic scenes, we propose a method of finding corresponding points through competition between dual correspondences. Then, we select point trajectories considered less corrupted by the errors, and use only these selected point trajectories for motion segmentation. Moreover, for motion segmentation in 3D scenes where strong perspective effects of 3D objects exist, we present a sampling method for generating motion hypotheses in 3D scenes.

Lastly, we handle the application of motion segmentation to generic object hypotheses generation for window-based object detection. We present a motion-based objectness score, which is computed using the motion-segmented points. Then, we combine the motion-based score with the existing appearance-based score, producing the combined objectness score of motion and appearance.

Through experiments, we demonstrated that the proposed motion segmentation algorithm rapidly ran online, producing the results comparable with those of recent offline methods, and also verified the effectiveness of proposed method in highly dynamic 3D scenes. Furthermore, by the use of the proposed combined objectness score, we could achieve superior results of generic object hypotheses generation.