Abstract

This paper presents an ideal method for object tracking directly in the compressed domain in video sequences. An enhanced rotation-invariant image operator called Largest Difference Indexed Local Ternary Pattern (LDILTP) has been proposed. The Local Ternary Pattern which worked very well in texture classification and face recognition is now extended for rotation invariant object tracking. Histogramming the LTP code makes the descriptor resistant to translation. The histogram intersection is used to find the similarity measure. This method is robust to noise and retain contrast details. The proposed scheme has been verified on various datasets and shows a commendable performance.