A Virtual Cinematographer for Presenter Tracking in 4K Lecture Videos

Lecture recording has become an important part of the provision of accessible tertiary education and having good autonomous recording and processing systems is necessary to make it feasible. In this work, we develop and evaluate a video processing framework that uses 4K video to track the lecturer and frame him/her in a way that simulates a human camera operator. We also investigate general issues pertaining to blackboard usage and its influence on cinematography decisions. We found that post-processing produced better tracking and framing results when compared to some real-time approaches. Furthermore, the entire pipeline can run on a commodity PC and will complete within the suggested time of 300% of the input video length. In fact, our testing showed that 60% of the total processing time can be ascribed to I/O operations. With the removal of redundant reads and writes, this proportion can be reduced. Finally, some algorithms can be remapped to parallel versions which will exploit multicore CPUs or GPUs if these are available