Fusion of Frequency and Spatial Domain Information for Motion Analysis

Abstract

This thesis investigates new approaches for the analysis of multiple motions in video, which integrates frequency and spatial-domain information. The tasks of interest are finding the number of moving objects, velocity estimation, object tracking, and motion segmentation. The proposed hybrid approach performs the motion estimation based on frequency-domain information, but also uses spatial information for precise object localization. Unlike existing frequency-domain methods, the use of this hybrid approach is not limited to constant translational motions, but can also address the problem of roto-translational and nonconstant motions. Frequency information is also used to detect and characterize multiple periodic motions in a video sequence. For this purpose, two methods using time-frequency distributions are presented. The first method is based on the time-frequency analysis of spatial projections of the video sequence, which is computationally efficient and leads to reliable results. The second method overcomes errors introduced by the projection method, by performing the analysis of the sequence in two dimensions. The resulting period estimates are then used to extract the periodically moving objects. The validity, effectiveness, and potential of all proposed approaches is verified through experiments with both synthetic and real video sequences.