Similar to last week, one of the main challenges to our system continues to be trying to simultaneously coordinate up to 5 live cameras, while maintaining acceptable performance in our CV detection system. As a solution, we have been working on migrating our video streaming to an online system, so that we can pull video and run our CV system on Google Colab instead. This way, we can leverage the Google cloud system to run our CV backend at a much higher speed than on our local CPUs. This week, we were able to setup our web streaming framework on the Raspberry Pis, and access the streamed video on the web servers in Google Colab.
We also continued to test our change from last week, a detection system comprised of 2 main cameras. Throughout this week, we experimented with partitioning estimation responsibilities of the various doorways across the 2 cameras. Between tests, we modified our detection logic and adjusted camera configurations in order to maximize our estimation accuracy. With a narrower range of focus for each camera in the 2 camera system, compared to trying to have 1 camera track all the doorways, we were able to simplify our detection logic, while improving our estimation consistency. A log of the various configurations and improvements we have made to our testing configuration can be found here: Test Log.
There is no significant changes to our development/testing schedule posted in our Team Status Report last week.