|
|
|
Poster Submissions
| |
| |
Multiresolution Segmentation for Video Surveillances
Mohammed A-Megeed Salem and Beate Meffert
Signal Processing and Pattern Recognition Group
Department of Computer Science
Humboldt-Universität zu Berlin, Germany |
|
Multiresolution analysis is a successive coarser and coarser approximations of the original signal. This is interpreted as representing the signal by different levels of resolution. The most obvious advantage of multiresolution representations is that they provide a possibility for reducing the computational cost of various image processing operations. Moreover they have the useful property of giving as well as global image features and local features. In particular they permit local interactions between features that are far apart in the original image where successive levels are constructed with lower resolution.
Since the theory for the multiresolution signal decomposition proposed by Mallat in 1989, the wavelet transform is the most commonly used method to implement the multiresolution transformation. The wavelet transform is a tool that cuts up data or functions or operators into different frequency component, and then studies each component with a resolution matched to its scale. That is, wavelets provide a tool for time-frequency localization. Transient features (short-time details) of a function f can easily be localized from looking at the wavelet coefficients, whereas longtime trends of f are stored in deeper layers of the coefficient hierarchy, as a consequence they are less precisely localized on the time axis.
We introduce a 3D wavelet-based segmentation algorithm for extracting moving objects in a traffic monitoring applications. The 3D wavelet transform gives the advantage of considering the relevant spatial as well as temporal information of the movement. A movement in an image sequence is a 3-dimensional change: in spatial-x, in spatial-y and in time. Many levels of analysis were considered. Fast motions are detected better in the first levels whereas slow motions or motion of big objects are detected better from the deeper layers. Therefore a cross-resolution combinations from the different levels were tested.
Experiments on different data sets, representing different traffic scenarios, show the robustness of the proposed algorithm against noisy images and the change of the lighting conditions. The results are compared with results obtained from a published 2D wavelet-based algorithm and they have been evaluated manually.
|
|
|
