Gabor Wavelet Feature Tracking Method

In this dissertation, a Gabor wavelet based method is used for tracking a target such as plane and vehicle with bad deformations including rotation, translation, scale and other affine transforms. The study of this method included feature abstraction; feature tracking and feature refreshing.2D continues Gabor wavelets and Levenberg-Marquardt (LM) algorithm are used to automatically abstract features by optimize Gabor wavelet's parameters of translation, orientation, and scale to make it approximates a local image contour region. The method of Sobel edge detection is used to get the initial position and orientation value of optimization in order to improve the convergence speed. The iteration numbers and feature abstraction effect are compared with the method in which the initial value is randomly created.The same LM algorithm is used to track the affine deformation parameters of the target. According to linearity between wavelet coefficient and its dual wavelet coefficient, the dual wavelet coefficient is directly used to save the computational time. An adaptive multi-resolution method is used to adaptively adjust the resolution of the object according to its size, thus the increase of computation caused by an extra-large target size is avoided.A feature refreshing method during tracking is studied. Greedy algorithm is used to compensate non-affine deformation caused by the changes of view and target's movement.Infrared and visible image sequences of vehicles and airplanes are used for tracking experiments under the situation of that the object is partially blocked by a momentary obstacle, or the illumination and contrast changes, or a big change of the scale and orientation of the object is happened. The result of the experiments shown that the method based on Gabor wavelet features can reliably track a target with affine transforms under bad situation mentioned above.