An Adaptive Semi-automatic Video Object Extration Algorithm Based on Joint Transform and Spatial Domains Features

We propose a new adaptive algorithm for semiautomatic video object segmentation based on joint pixel features using the undecimated wavelet packet transform (UWPT) and luminance value. The method starts with the object’s boundary specification at the reference frame assisted by the user. After selecting a set of feature points which approximate the object’s boundary, the amplitude of coefficients in the best basis tree expansion of the UWPT is used to create a Feature Vector (FV) corresponding to each pixel. Weighting the FV with the magnitude of the pixel’s luminance results in a pixel-wise feature that can be tracked temporally in the video sequence. Full search for the best match has been performed through the use of generated FVs and an adaptive search window updating. Experimental results show a good performance in case of object translation, small rotation and scaling. The method can be used to track both rigid and non-rigid shapes in video and image sequences.