Towards active event recognition

摘要: Directing robot attention to recognise activities and anticipate events like goal-directed actions is a crucial skill for human-robot interaction. Unfortunately, issues intrinsic time constraints, the spatially distributed nature of entailed information sources, existence multitude unobservable states affecting system, latent intentions, have long rendered achievement such skills rather elusive goal. The problem tests limits current control systems. It requires an integrated solution tracking, exploration recognition, which traditionally been seen as separate problems in active vision. We propose probabilistic generative framework based on gain maximisation mixture Kalman Filters that uses predictions both recognition attention-control. This can efficiently use observations one element dynamic environment provide other elements, consequently enables guided exploration. Interestingly, sensors policy, directly derived from first principles, represents intuitive trade-off between finding most discriminative clues maintaining overall awareness. Experiments simulated humanoid observing human executing goal-oriented demonstrated improvement precision over baseline