“…Feature-level fusion retains less detailed information than data-level fusion, but it is also less prone to noise and sensor failures, and, most importantly, it is the most appropriate type of fusion for tightly coupled and synchronized modalities. Though many feature-level techniques like Kalman fusion, artificial neural networks (ANN) based fusion, and hidden Markov models (HMM) based fusion have been proposed [26], [105], decision-level (i.e., interpretation-level) fusion is the approach applied most often for multimodal HCI [65], [88], [105]. This practice may follow from experimental studies that that have shown that a late integration approach (i.e., a decision-level fusion) might provide higher recognition scores than an early integration approach (e.g., [100]).…”