Human Behaviour Visualisation and Simulation for Automatic Video Understanding

The objective of this work is the visualisation and simulation for automatic video interpretation. We have conceived a test framework that generates 3D animations corresponding to behaviours recognised by an automatic interpretation system or corresponding to behaviours described by an expert. Conceiving this test framework is essential in order to be able to develop and validate the interpretation process. The objective of our test framework is (1) to visualise the computation of the interpretation, (2) to be flexible (configurable) enough for testing the different configurations of the interpretation and (3) to be realist enough to understand what is interpreted. To solve this problem we have defined six types of model to represent all the information that is necessary for the interpretation. First, we propose a model of the scene context (containing the 3D geometry) and a model for the virtual camera. Second, we propose an articulated and hierarchical model for representing the human body given its sub parts. We propose two other hierarchical models for modelling human actions and scenarios, and also a model of scene-scenarios that gathers all previous models. We have defined a description language for representing these models. The obtained results are promising: we have developed a test system for a given interpretation system and started evaluating it by generating test animations.