Improving human action recognition using decision level fusion of classifiers trained with depth and inertial data

Official URL: http://risc01.sabanciuniv.edu/record=b1817059 (Table of Contents)

Improvement in sensor technology has aided research in the field of human action recognition (HAR), as acquiring data is easier and the obtained data is more accurate. However, each sensor has its own limitations and benefits, and a combination of these sensors can help improve the accuracy of recognition systems. This thesis presents an in depth study of HAR using decision level Improvement in sensor technology has aided research in the field of human action recognition (HAR), as acquiring data is easier and the obtained data is more accu-rate. However, each sensor has its own limitations and benefits, and a combination of these sensors can help improve the accuracy of recognition systems. This thesis presents an in depth study of HAR using decision level fusion of classifiers that are trained using RGB-D camera and inertial sensor data. Extraction of ro-bust and subject-invariant features is performed to train independent classifiers, i.e. neural networks, for action recognition purposes. This work employs decision level fusion on the outputs of the individual classifiers using a probabilistic approach in the form of Logarithmic Opinion Pool (LOP). The effect of varying the parameters of the proposed algorithm on the final 8-fold cross-validation accuracy is analyzed. The proposed algorithm is tested on UTD-Multimodal Human Action Dataset that contains actions which are based upon the movement of different set of joints, and it achieves an average 8-fold cross-validation accuracy of 97.3%.