Next: Approaches Up: Introduction Previous: Control


A further development in this subfield involves representations to support task-level control and learning. For example, Hidden Markov Models (HMMs) were used for learning probabilistic relationships for eye movement control [51] and applied to modelling of vehicle trajectories [26]. On-line updating using the visually augmented HMMs then allowed both tracking and reporting of these purposive vehicle movements. More recently, using Bayesian Belief Networks (BBNs) supported the learning of both initial and conditional probabilities for camera control [52] and for segmenting and tracking vehicles [27]. In addition, BBNs have been used with behavioural models to provide task-dependent control in behavioural analysis [15,33]. These kinds of learning are essentially conditional parameter estimation using the statistics of example image sequences. Learning dynamic, parametric models for visual motion patterns [7] is an important capability for intelligent tracking. Also, learning of statistically-based deformable models is crucial for many medical applications [17], and in tracking moving people [3] for surveillance. In these examples, the knowledge is acquired off-line and exploited in the on-line system. We can extend the role of learning for behavioural modelling using on-line evaluation or reinforcement learning [53,69] to create more open systems that adapt their own behaviour to the environment. This is an exciting field in which we can envisage fully autonomous visual agents learning their own goals and representations.