Ask HN: Offline forecasting model to real-time sensor data?

I have a model that predicts crowd density at transit stations using months of historical turnstile data (node + flow features). Works great offline. Now I want the same thing from real-time video — person detections aggregated into zone counts every second. No historical corpus, noisy signal, much shorter time scale. Pre-train on structured data and transfer? Build a simpler online model? Any pointers? Thank you