Action recognition by graph embedding and temporal classifiers
- Publication Type:
- Issue Date:
With the improved accessibility to an exploding amount of video data and growing demand in a wide range of video analysis applications, video-based action recognition becomes an increasingly important task in computer vision. Unlike most approaches in the literature which rely on bag-of-feature methods that typically ignore the structural information in the data, in this monograph we incorporate the spatial relationship and the time stamps in the data in the recognition and classification processes. We capture the spatial relationships in the subject performing the action by representing the actor’s shape in each frame with a graph. This graph is then transformed into a vector of real numbers by means of prototype-based graph embedding. Finally, the temporal structure between these vectors is captured by means of sequential classifiers. The experimental results on a well-known action dataset (KTH) show that, although the proposed method does not achieve accuracy comparable to that of the best existing approaches, these embedded graphs are capable of describing the deformable human shape and its evolution over time. We later propose an extended hidden Markov model, called the hidden Markov model for multiple, irregular observations (HMM-MIO), capable of fusing spatial information provided by graph embedding and the textural information of STIP descriptors. Experimental results show that recognition accuracy can be significantly improved by combining the spatio-temporal features with the structural information obtaining higher accuracy than from either separately. Furthermore, HMM-MIO is applied to the task of joint action segmentation and classification over a concatenated version of the KTH action dataset and the challenging CMU multi-modal activity dataset. The achieved accuracies proved comparable to or higher than state-of-the-art approaches and show the usefulness of the proposed model also for this task. The next and most remarkable contribution of this dissertation is the creation of a novel framework for selecting a set of prototypes from a labelled graph set taking class discrimination into account. Experimental results show that such a discriminative prototype selection framework can achieve superior results, not only for the task of human action recognition, but also in the classification of various structured data such as letters, digits, drawings, fingerprints compared to other well-established prototype selection approaches. Lastly, we change our focus from the forementioned problems to the recognition of complex event, which is a recent area of computer vision expanding the traditional boundaries of visual recognition. For this task, we have employed the notion of concept as an alternative intermediate representation with the aim of improving event recognition. We model an event by a hidden conditional random field and we learn its parameters by a latent structural SVM approach. Experimental results over video clips from the challenging TRECVID MED 2011 and MED 2012 datasets show that the proposed approach achieves a significant improvement in average precision at a parity of features and concepts.
Please use this identifier to cite or link to this item: