Anyway, despite the camera's strengths, there are a few limitations. First, the camera can only support 4 simultaneous clients and the performance degrades linearly with each additional client (from my anecdotal experience). Second, the only access control the camera offers is HTTP Basic Authentication backed with a 4-user list configurable from its web interface that doesn't integrate well with any other application or security system. I figured that the best and most direct way of fixing the problems was to proxy the feed through my MacPro to manage the connection and user access there instead of on the camera.
As I was imagineering this system, I also got the bright idea to go ahead and do facial detection (not recognition -yet-) on the stream. After doing some research on the technology, I decided to use the OpenCV libraries developed by Intel and subsequently open sourced. My initial prototypes were extremely slow (1-2 FPS) since the Java libraries depended on JNI calls to a non-thread-safe C library. I did more research and found the Faint (Face Annotation Interface) library which did Haar in pure, multithread-able Java. (I had to take the beta code from the SVN since it wasn't released yet.) That finally got me a much more acceptable 10+ FPS.
Now I have the camera stream being cached within a custom-built Tomcat webapp that does the detection and also provides security for the stream. It can support much more than the 4 users available from the camera and without a FPS hit. It's pretty cool. Right now it just draws a red rectangle around the detected face, but obviously more triggers and actions are possible and desirable. It should definitely be noted that the stream (with facial detection) is viewable from the iPhone! Now- to just get the damned thing attached to my iRobot and my little mobile sentry will be complete. :)