Output details

Structure Inference for Bayesian Multisensory Scene Understanding

<22>Originality: Presents one of the first uses of Bayesian structure inference to solve multi-sensor cue integration, extending treatment of the problem (for the first time) beyond pure sensor fusion and maximum likelihood methods.

Significance: Provides the first practical implementation of robust audio-visual tracking where the multimodal cue source is intermittent. Framework has spawned several novel interpretations of human sensorimotor adaptation paradigms that allow incorporation of motor variability and quality of learned priors.

Rigour: The paper provides complete derivation and proofs of all update equations. Hardware experiments illustrate real-time performance, including robust detection of video occlusions, sound source disambiguation and tracking.