To make this work, you need to activate multi ControlNet in Settings -> ControlNet -> Multi ControlNet: Max models amount. I have this set to 4. Requires a restart.

Workflow

Drop the background OpenPose file into ControlNet Unit 0, Control Weight 0.5. Rest stays at default.
Drop the depth map into ControlNet Unit 1, Control Weight 0.35 (may need to be tweaked some; sometimes 0.3 works better). Rest stays at default.
Drop your foreground character's openpose file into ControlNet Unit 2, everything, including Control Weight, stays at default. This can be any pose of your choice. Ideally the aspect ratio matches the one of the 2 background files (they're all 3x2 or 2x3).
Obviously you need to set pre-processor to "none" for all of the 3 ControlNet units.
Useful key words for the prompt: "crowd, multiple_others" ("multiple_boys", "multiple_girls" if it applies). If you're generating a stadium or arena, you can add "spectators, arena, stadium". Or theater, courtroom, concert, etc. Generating scenes with crowded streets has also worked.
This looks like it might work best with depth of field (?).
Prompting at higher resolutions might generate more background character than going 768x512 or 512x768 (?). Still unsure about this. Many generations were just fine at low resolutions.