The radiosity principle for the simulation of global illumination in virtual scenes, which is now known for 13 years, has been developed and enhanced several times since then, but it still requires intensive interaction from the user and the actually necessary parameters for the entire radiosity process haven't been clearly defined down to the present day.

This Diploma thesis is addressing the various aspects composing a radiosity simulation and creates a full concept for a hierarchical radiosity technique, which uncovers specific problems of existing techniques and offers corresponding solutions.

One of the main reasons that the radiosity method fails to play an important role in computer graphics so far is the immense number of intertwined parameters, which can be adjusted correctly only by an experienced user with distinct knowledge of the radiosity principle. Part of this thesis is therefore dedicated to find automatisms for these parameters and to present default values and strategies for those for which a full automation is not advisable.

The scene subdivision during the radiosity computation is generally performed by a refinement criterion in combination with an adaptive meshing technique. This thesis is investigating in which cases a subdivision is actually necessary. A combined refinement criterion is devised which guarantees a fixed error bound for the radiosity solution, but which is also based on perceptive aspects to suppress a subdivision where the current mesh is already of sufficient quality.

Finally, a new method for visibility determination is developed which is not based on the faulty ray casting method and guarantees that none of the shadow borders are missed.

The described concept has been implemented and integrated into the Genesis2 system. The underlying hierarchical radiosity method now allows to create converged radiosity solutions within managable times, while the enhanced visibility algorithm together with the perceptual subdivision criterion creates a high quality result while keeping the number of resulting patches as small as possible.