Multiple kernel learning for first-person activity recognition

Özkan, Fatih
First-person vision applications have recently gained increasing popularity because of advances in wearable camera technologies. In the literature, existing descriptors have been adapted to the first-person videos or new descriptors have been proposed. These descriptors have been used in a single-kernel method which ignores the relative importance of each descriptor. On the other hand, first-person videos have different characteristics as compared to third-person videos which are captured by static cameras. Throughout the first-person video, vast changes occur in some attributes such as illumination or brightness. A significant amount of ego-motion is created because of the movements of the first-person camera wearer. Multiple features are used in order to capture the different changes in video characteristics. Therefore, appropriate feature and kernel selection are needed. In this thesis, local and global motion-related features are used. A data-driven approach is proposed in order to select and combine these features and kernels employed. Feature and kernel selection is performed through AdaBoost algorithm’s well-known trials in a probabilistic manner. At training stage, a classifier which shows better performance than other classifiers is determined for each trial. After all trials, classifiers which compose the final classifier are determined. At testing stage, final classifier makes decision for activity labels based on a voting mechanism. Experiments show that the proposed methods outperform the traditional SVM single kernel-based methods in literature in terms of recognition accuracy. 
Citation Formats
F. Özkan, “Multiple kernel learning for first-person activity recognition,” M.S. - Master of Science, Middle East Technical University, 2017.