After testing the opencv samples for a bit I realized that a filtered image stream can be very handy in saving processing and storage. The idea is to utlize gstreamer chaining to slot in an openCV facedetect module and save only frames where face is detected.
I can currently capture timestamped frames using gstreamer:
gst-launch v4l2src num-buffers=1 ! video/x-raw-yuv,width=640,height=480,framerate=30/1 ! ffmpegcolorspace ! jpegenc ! filesink location=$(date +"%s").jpg
Then I can use OpenCV peopledetect and facedetect in the chain with queued up frames and start saving once the detection starts producing output frames. This will greatly reduce the amount of boring hallway frames and the board will concentrate on processing interesting events i.e. people approaching the camera. This can then be used for interesting data mining work such as time spent in the office, walking speed, hallway meetings between colleagues and number of days without changing shirts.
Time to ensure the gstreamer opencv module works properly on beagleboard.