ARTIFICIAL BROADCASTING · REAL-TIME SYSTEM ARCHITECTURE

One camera. Several virtual operators. One synthetic director.

A real-time broadcasting system that converts one wide-angle scene into multiple candidate views and decides which view should be shown next.

The prototype was built around a deliberately difficult constraint: a single centrally placed wide-angle camera, uncertain camera parameters, predefined output-screen shapes and no human camera operator. The system had to follow the action, avoid missing important events and still produce a visually useful broadcast.

Watch demo ↗Original article ↗

Artificial Broadcasting prototype — real-time state with wide scene, detected regions and generated operator views — Working prototype: the wide scene, active regions, generated views and internal decision state are visible at the same time.

Artificial Broadcasting prototype — another real-time state — Different game state → different candidate views.

Artificial Broadcasting prototype — another real-time state with changed virtual camera views — The output changes as the scene changes.

The task

Build an automatic broadcast, not merely an object tracker.

The system had to transform one physical camera stream into something closer to a small production team. It needed to detect local situations, form useful close-up or overview views, and decide which of those views best matched the current event.

physical scene→local situations→candidate views→editorial decision→broadcast

Three-level architecture

The core design was hierarchical. Each layer solved a different problem and passed a more abstract representation upward instead of forcing one algorithm to understand the whole broadcast.

LEVEL 01Situation detectors

Motion, position change, position stability and related low-level mechanisms identify local objects and events worth observing.

LEVEL 02Virtual operators

Several synthetic camera operators convert detector outputs into candidate views with different framing, shape and zoom.

LEVEL 03Synthetic director

The director compares the competing operator streams together with heuristics about spectacle and game strategy, then selects or combines the views for the output screens.

Why it mattered

The difficult part was not detecting motion. It was deciding what representation the next level of the system should receive.

The project became an experiment in hierarchical intelligence: local algorithms operate on image-level evidence, virtual operators operate on scenes, and the director operates on competing interpretations of the event.

Abstraction by levelEach layer works with a representation appropriate to its task instead of inheriting all low-level image complexity.

Competing viewsSeveral valid candidate views can exist simultaneously. A higher-level process decides which one currently serves the broadcast.

Real-time coordinationThe useful result comes from coordination between specialized modules, not from one monolithic detector.

Human roles as an engineering clueCamera operator and director roles were used as a functional decomposition of the problem, not as a requirement to imitate people literally.

The same physical scene is progressively transformed: pixels become situations, situations become views, and views become an editorial decision.