I've been developing auto lecture recording system since 2008. As a key component of the system, I have implemented a prototype of virtual camerawork(post-processing camerawork) generator. In our setting, we record the entire classroom with a fixed high-definition(1920x1080) camcorder. Then, generate a camera work that traces region of interest. Finally, a trimmed
version (720x480) of video is generated.
With a high-definition camcorder, we can record whiteboard and projector screen simultaneously at a sufficient resolution. In addition, with a HDD-recording model, we don't need to use expensive video capture board for high-definition video; what we need is just copying AVCHD video clips to PC.
One drawback of this method lies in CPU intensive work of AVCHD decoding. Under my current implementation (Mac Mini, Core 2 Duo 2GHz), it takes 6 hours to process a 90-minute lecture video. However, this work can be accelerated by applying hardware decoders on recent video cards (Is somebody already hacking VDPAU ?).
Example video clips:
These are example short clips from my lecture.
1. original video (H.264, HTTP Port 80 Streaming)
rtsp://atlantis.cc.kumamoto-u.ac.jp/DSA-original-95250-40.mp4
2. processed video (H.264, HTTP Port 80 Streaming)
rtsp://atlantis.cc.kumamoto-u.ac.jp/DSA-trimmed-95250-40.mp4
First one is the original video (though it is resized to save network bandwidth for streaming). Second one is the trimmed version generated by our virtual camerawork.As you see in this video clip, by jumping virtual camera position, we have a camera switch effect as well as pan motions. The video titles were automatically attached by detecting the QRcodes on the projector screen. (QRcode is 2d barcode commonly used in Japan).
For details, please see the attached paper and presentation slides.
Reference:
Takayuki NAGAI,
"Simple lecture recording with HDV camera and virtual camerawork", Proceedings of the 8th International Conference on Applications and Principles of Information Science,pp.321-324,2009.
