This paper presents a novel approach to the extraction of video features for real-time detection of neonatal seizures. In particular, after identification of a proper Region Of Interest (ROI) within the video frame, the broadening factor and the maximum distance between consecutive pairs of zeros of a properly extracted average differential luminosity signal are shown to be relevant features for a diagnosis. The ROI is selected by defining an area around the point where the maximum amplitude of the optical flow vector of that video frame sequence is observed. The located point is then tracked by an algorithm based on template matching and optical flow. The proposed approach allows to differentiate pathological movements (e.g., clonic and myoclonic seizures) from random ones.