Czech Technical University in Prague Doctoral Thesis Statement

A new salient semantics-less blob detector named the Stable Wave Detector (SWD) is suggested. The detector proved to be particularly very useful in tracking a target in a videosequence. SWD is a multiscale blob detector in the intensity image. The targets are blobs in 2D which correspond to local maxima/minima of the image intensities or positive and negative peaks in 1D. SWD belongs to a group of interest point/regions-like operators aiming at detecting repeatedly distinguished entities regardless of their meaning. Precision, robustness and speed of the SWD detector are main issues. The detector is based on the phase of the first harmonic wave in the window sliding across the image. Its localization is a result of an integral transformation rather than a usual derivative. Thus, the blob detector is inherently robust to noise and blur in the image. The SWD provides a subpixel localization of blobs together with the estimate of its precision, the measure of the strength/significance and the estimate of the size/scale for each blob. Detected blobs have some minimal distance given by the period used for detection. Their minimal mutual distance can facilitate matching of corresponding scene points. The properties of the SWD detector itself did not beat the state of the art considerably. The situation has improved dramatically when the SWD detector was employed in tracking a target, i.e. in establishing a point to point (blob to blob) correspondences across the videosequence. The outcome is an ultrafast tracker able perform better than the stateof-the-art. The tracker takes two inputs: (a) the location of the blob in the previous frame estimating its location in the current frame and (2) the current frame in which the blob should be found. The problem of tracking is considered in several levels of abstraction. At a low level, correspondences between individual points are established at different scales independently. The high level of tracking system incorporates a model of the movement and a model of the target. The trade off between the precision and the range of trackable displacement at different scales must be solved. It is shown how the movement and target model can be used in a coarse to fine approach solving the trade off between the precision and the range. The SWD detector and the tracker utilizing it were extensively tested experimentally. Test results are presented in the thesis. The implementation of the involved algorithms in C language is provided to the research community.