In recent years, action recognition based on RGB-D data has attracted increasing attention. Different from traditional 2D action recognition, RGB-D data contains extra depth and skeleton modalities. Different modalities have their own characteristics. This thesis presents seven novel methods to take advantages of the three modalities for action recognition. First, effective handcrafted features are designed and frequent pattern mining method is employed to mine the most discriminative, representative and nonredundant features for skeleton-based action recognition. Second, to take advantages of powerful Convolutional Neural Networks (ConvNets), it is proposed to represent spatio-temporal information carried in 3D skeleton sequences in three 2D images by encoding the joint trajectories and their dynamics into color distribution in the images, and ConvNets are adopted to learn the discriminative features for human action recognition. Third, for depth-based action recognition, three strategies of data augmentation are proposed to apply ConvNets to small training datasets. Forth, to take full advantage of the 3D structural information offered in the depth modality and its being insensitive to illumination variations, three simple, compact yet effective images-based representations are proposed and ConvNets are adopted for feature extraction and classification. However, both of previous two methods are sensitive to noise and could not differentiate well fine-grained actions. Fifth, it is proposed to represent a depth map sequence into three pairs of structured dynamic images at body, part and joint levels respectively through bidirectional rank pooling to deal with the issue. The structured dynamic image preserves the spatial-temporal information, enhances the structure information across both body parts/joints and different temporal scales, and takes advantages of ConvNets for action recognition. Sixth, it is proposed to extract and use scene flow for action recognition from RGB and depth data. Last, to exploit the joint information in multi-modal features arising from heterogeneous sources (RGB, depth), it is proposed to cooperatively train a single ConvNet (referred to as c-ConvNet) on both RGB features and depth features, and deeply aggregate the two modalities to achieve robust action recognition.
History
Year
2017
Thesis type
Doctoral thesis
Faculty/School
School of Computing and Information Technology
Language
English
Disclaimer
Unless otherwise indicated, the views expressed in this thesis are those of the author and do not necessarily represent the views of the University of Wollongong.