Quantifying motion in video recordings of neonatal seizures by feature trackers based on predictive block matching

Abstract

This work introduces predictive block matching, a method developed to track motion in video by exploiting the advantages of block motion estimation and adaptive block matching. The proposed method relies on a pure translation motion model to estimate the displacement of a block between two successive video frames before initiating the search for the best match of the block tracked throughout the frame sequence. The search for the best match relies on adaptive block matching, which employs an update strategy based on Kalman filtering to account for the changing appearance of the block. Predictive block matching was used to extract motor activity signals from video recordings of neonatal seizures.