Continuous wavelet transform for time-varying motion extraction

The widespread use of digital multimedia data has made the development of advanced processing techniques necessary, to enable its more efficient analysis. For video content, the estimation of motion is a fundamental step in the extraction of activity, for tracking, motion segmentation, video classification and other applications. The numerous methods that have been proposed over the years for the problem of motion estimation can be divided into two categories. The first group processes data in the spatial domain, and the other in the frequency domain. In this work, an original approach for the estimation of motion in the frequency domain is presented. The proposed method avoids limitations of illumination-based methods, such as sensitivity to local illumination variations and noise by employing the continuous wavelet transform (CWT). All video frames are processed simultaneously, so as to create a frequency-modulated (FM) signal, which contains the motion information in its frequency. The resulting FM signal is then processed using the CWT, which extracts its time-varying frequency and consequently its motion. This system is shown to be robust to local measurement noise and occlusions, as it processes the available data in a global, integrated manner. Experiments take place with both synthetic and real video sequences to demonstrate the capabilities of the proposed approach.