Advanced Deep Learning Network with Harris Corner based Background Motion Modeling for Motion Tracking of Targets in Ultrasound Images

We have proposed a novel background motion estimation method to improve the tracking of targets in advanced Siamese networks. One issue with ST is that it does not adapt to video-specific cues. The main problem, however, is that no motion of the object is assumed, and the last position is used as the center of the search region for the next frame. We propose to accurately model the motion of the object of interest by accounting for the background motion present in the frame. A novel method for estimating the background motion based on the Harris corner detector is proposed due to its robust feature-point selection. We further adopted a more recent version of ST, Correlation Filter Network (CFNet) which uses an adaptive Correlation Filter Layer (CFL) to efficiently learn the cues present in the video. The results obtained on the Carotid Artery (CA) dataset demonstrate that the proposed method outperforms other similar tracking approaches.