NAOS: NAtural language description of Object movements in Street scenes

Automatic Event Recognition of Agents in Street Scenes

The software was developed at the University of Hamburg by Bernd Neumann, Klaus Strobel and Rainer Anscheit, 1984-1988.

The NAOS contains a 2D trajectory editor for specifying a spatiotemporal scene with multiple agents, a 3D scene color animation viewer, and a recognition component for events in the scene. The so-called geometric scene description (GSD), i.e. a symbolic description of a scene provides the basis for event recognition. The NAOS system is oriented towards recognizing events in street scenes.

The color pictures below were rendered on a 3640 machine and show a model of a predefined 3D street scene in Hamburg near the Computer Science Department at that time (Schlüterstraße, Bieberstraße, Hartungstraße). Animation was done by moving color bitmaps in the 3D scene. Unfortunately, no pictures are available of the hardware configuration.

The buildings resemble the actual buildings in that street. The overall motivation behind the 3D model was to acquire a 3D model with Computer Vision techniques (see the taxi pictures from an image sequence below) and to describe occurring event with a natural language module. The NAOS street model was hand-crafted, though. Spatiotemporal trajectories of agents (pedestrians, cars) can be specified within the static street scene using the Trajectory Editor software written with Genera. The 3D animator was used to check the plausibility of the input to the NAOS event recognition system and to verify its output.

Experiments with a learning component for typical trajectories for certain events (e.g. turning off) have been carried out. The non-propositional representation structure was called (TAF, trajectory accumulation frame).