The analysis of visual motion against dense background clutter is a challenging problem. Uncertainty in the positions of visually sensed features and ambiguity of feature correspondence call for a probabilistic treatment, capable of maintaining not simply a single estimate of position and shape, but an entire distribution. Exact representation of the evolving distribution is possible when the distributions are Gaussian, and this yields some powerful approaches. However, normal distributions are limited when clutter is present: because of their unimodality, they cannot be used to represent simultaneous alternative hypotheses. One powerful methodology for maintaining non-Gaussian distributions is based on random sampling techniques. The effectiveness of 'factored sampling' and 'Markov chain Monte Carlo' for interpretation of static images is widely accepted. More recently, factored sampling has been combined with learned dynamical models to propagate probability distributions for object position and shape. Progress in several areas is reported here. First a new observational model is described that takes object opacity into account. Secondly, complex shape models to represent combined rigid and non-rigid motion have been developed, together with a new algorithm to decompose rigid from non-rigid. Lastly, more powerful dynamical prior models have been constructed by appending suitable discrete labels to a continuous system state; this may also have applications to gesture recognition. [References: 48]
展开▼
