Perceptually-Adaptive Collision Detection for Real-time Computer Animation

Perceptually-Adaptive Collision Detection for Real-time Computer Animation Carol O'Sullivan Supervisor: Dr. Steven Collins The aim of interactive animation systems is to create an exciting and real experience for viewers, to give them a feeling of immersion, of “being there". The tendency in the past has been to attempt to achieve this by matching as closely as possible the physics of the real world, with varying degrees of success. However, it is the human visual system that receives and interprets the visual cues from the surrounding environment, and it ultimately determines what we perceive. Therefore, we must look beyond the laws of physics to find the secret of reproducing visual reality. In interactive animation applications such as VR or games, it cannot be predicted in advance how a user or the entities in a virtual world will behave, so the animation must be created in real-time. There are many bottlenecks in such systems, collision detection being a major one. A trade-off between detection accuracy and speed is necessary to achieve a high and constant frame-rate. However, it is possible to reduce perceived inaccuracy by taking perceptual factors into account, and also by estimating where on the screen a viewer is looking, possibly using an eye-tracking device, or by attaching more importance to certain objects in a scene, or to regions of the screen. In this thesis we present the first perceptually-adaptive collision detection algorithm. New collision scheduling strategies are also presented and evaluated, along with a new interruptible algorithm to test for the intersection between two sphere trees. A model of human visual perception of collisions is developed, based on twodimensional measures of eccentricity and separation. The model is validated by performing psychophysical experiments, in the first study of its kind. We demonstrate the feasibility of using this model as the basis for perceptual scheduling of interruptible collision detection in a real-time animation of large numbers of homogeneous objects. The user's point of fixation may be either tracked or estimated. By using a priority queue scheduling algorithm, perceived collision inaccuracy was significantly improved. The ideas presented here are applicable to other tasks where the processing of fine detail leads to a computational bottleneck.