Python port of Zana Zakaryaie's FAST-KLT tracker originally written in C++. I'm still working on the implementation. This is the tracker's current output:
trackerOut.mp4
The algorithm first uses a caffe deep model for the initial face detection. It then extracts FAST keypoints from this ROI and tracks them using KLT. The new keypoints are used to calculate a new bounding rectangle on the new frame. The result is one initial detection and a fast, smooth (and somewhat robust to occlusion) tracking for new frames. Full details here.