The goal is to achieve as close to green screen results as possible masking out people's silhouettes as they walk by, but in varied installations environments. The ideal is to find options that works in every lighting condition, with noisy changing backgrounds, in realtime and with multi camera support. We've tested a variety of approaches including machine learning + RGB camera, machine learning + Depth camera, and various depth cameras with 3D masking.
Also see our Skeleton Detection Research
More information and notes on our github.