Recognition of neonatal seizures from video recordings based on motion tracking methods

This dissertation presents several motion tracking methods developed to quantify motion information from video recordings of neonatal seizures in the form of temporal motion trajectory signals. The motion tracking methods were developed relying on a variety of block motion models, which include a translation model, an affine model, a fractional model, and a generalized fractional model, and by minimizing different tracking error functions. Those motion trackers were further developed to adjust to illumination and contrast changes in video recordings. The quantitative features that convey some unique behavioral characteristics of neonatal seizures are extracted from the motion trajectory signals produced by different motion tracking methods. A learning algorithm is proposed for training cosine radial basis function neural networks capable of identifying uncertainty in the classification of multidimensional data. The motion tracking methods developed in this study were evaluated based on the performance of different kinds of neural network models including conventional feed-forward neural networks, quantum neural networks, cosine radial basis functions neural networks trained by the original learning algorithm and cosine radial basis functions neural networks. These models were trained by the proposed learning algorithm, trained to recognize the neonatal seizures using a set of 240 video recordings. The experiments indicated that the motion tracking methods developed in this dissertation produced quantitative features that constitute a reliable basis for detecting neonatal seizures.