Cross-View People Tracking by Scene-Centered Spatio-Temporal Parsing