Kalman Filter Based Classifier Fusion for Affective State Recognition

摘要: The combination of classifier decisions is a common approach to improve classification performance [1–3]. However, non-stationary fusion still research topic which draws only marginal attention, although more and systems are deployed in real-time applications. Within this work, we study Kalman filters [4] as combiner for temporally ordered decisions. filter linear dynamical system based on Markov model. It capable combining variable number measurements (decisions), can also deal with sensor failures unified framework. analyzed the setting multi-modal emotion recognition using data from audio/visual emotional challenge 2011 [5, 6]. shown that well-suited fusion. Combining available sequential uni- does not result consistent continuous stream decisions, but leads significant improvements compared input decision performance.