Human Action Recognition Using Fusion Of Depth And Inertial Sensors

In this paper we present a human action recognition system that utilizes the fusion of depth and inertial sensor measurements. Robust depth and inertial signal features, that are subject-invariant, are used to train independent Neural Networks, and later decision level fusion is employed using a probabilistic framework in the form of Logarithmic Opinion Pool. The system is evaluated using UTD-Multimodal Human Action Dataset, and we achieve 95% accuracy in 8-fold cross-validation, which is not only higher than using each sensor separately, but is also better than the best accuracy obtained on the mentioned dataset by 3.5%.