Application of the gradient orientation for systems of automatic target detection

In this paper, a problem of automatic target tracking in complex natural backgrounds is considered. Target detection is performed in each frame of a video sequence by the elementwise comparison with a reference image. The proposed method is based on the representation of every pixel as the orientation of the luminance gradient in the vicinity. The vicinities are divided into classes depending on their orientation. In addition to the classes of anisotropic vicinities, a class of vicinities with an isotropic structure is introduced. The classes are numbered and the number of the vicinity class is used as a feature of the point of interest. Thus, the original gray-scale image is transformed to a pseudo-image in which the detection procedure is carried out. The encoded image is then scanned using a reference image. The elementwise comparison of the reference image with the current fragment is performed in a feature space. As a result, a comparison matrix is formed, each element of which is the number of matching elements of the reference image and the current image fragment. The position of the reference image is determined by the maximum value of the comparison matrix. A special rule of reference image overwriting, the so-called dynamic proximity measure, is used to achieve stable tracking. The testing results have shown that with our approach the object tracking is more stable in comparison with the use of normalized correlation.