Those options can improve performance, specially the step size and initial scale. ViolaJones is a good object detection algorithm, although for large images 640x480 it can reach low frame rates.
Two features from our roadmap that will speed up performance for cases like that:
1- Create some sort of ThreadableTracker abstract class that splits process into different Workers.
2- Use optical flow to predict the movement in order to not require processing every frame using ViolaJones.
A workaround for now is to write the 640x480 video into a smaller hidden canvas, and this canvas is the element being track, when track event fires with rectangles you can proportionally projects the coordinates into the larger visible video.