Automated cinematography for games.

This thesis deals with the issue of automated cinematography for games. In 3D videogames, the system must continuously provide the player with a view of the virtual world and its characters. The difficulty is that contrary to the cinema the actors are unpredictable. In particular the player continuously modifies the virtual environment by moving objects or by interacting with the other non-playable characters. The latter, because of their more and more sophisticated artificial intelligence, can have behaviours that were not predicted by the developers themselves (such as the complex behaviours that emerge from the combination of basic behaviours).
Some games have solved the problem by predefining the possible positions of the camera during the game development while some others give control of the camera system to the player, so that he can find by himself the best possible view. I aim however at finding an intermediate solution, where the camera system would automatically generate both engaging and usable views. The camera system should be able to adapt to every situation of the virtual world without user intervention, and should allow the player to interact with his surrounding in the most efficient way. Such a camera system could be of interest for the game industry. Currently, in many games, the camera movements, positions, etc. are set using scripts manually written by the developers. Having a fully automated system could potentially save hours of work. This system could also be used for the 3D virtual worlds or “3D chats” on the Internet. For example, the avatars – the characters played by the users – could be “filmed” in a different way depending on the mood of the users. I aim to develop techniques which can be generalised to these and other areas of application.
Existing approaches to automated cinematography will be reviewed – focusing on the constraint-based and idiom-based ones – in order to highlight the strength and limitations of each one. A solution to the problems found will be proposed in the form of a camera system implemented using Adobe Director. It will be based on “rules” derived from existing cinematographic knowledge. One of my aims will also be to show that using generic rules can give results close to the idiom-based approaches with the convenience of being able to adapt to any type of scene.