One of the more interesting pieces of open source software for video processing is OpenPose. I used this code as the basis of a new OpenPoseGPU Stream Processing Element for rt-ai and the results can be seen in the screen capture. The metadata produce can be seen partially on the right hand side – it is pretty extensive as it contains all of the detected key points, depending on whether face and hand processing is enabled.
This version is x86/NVIDIA GPU based. The next thing to do is to get the equivalent working with the Intel NCS 2, based on this example, and then compare performance to see if the NCS 2 is practical for applications needing specific frame rates. The goal is to generate metadata that can be used to train a deep neural net to recognize specific activities. This could be used to create a Stream Processing Network that generates high level metadata about what users in the view of the camera are doing. This is in turn could be used to generate feedback to users, generate alerts on anomalous behavior etc.