Research On Contrast Pattern Mining Based Human Action Recognition

Studies on human action recognition bridge the gap between raw data (videos or images) and high-level semantics of human by analyzing the raw data and extracting valuable information from the raw data for the actions involved in. This research topic gives raise to many applications including intelligent video surveillance, human-computer interaction, video indexing/retrieval and sport video anlaysis. As a new research subject in visual understanding domain, human action recognition is built on the achievements of many existing research subjects, such as computer vision, pattern recognition and cognitive science. Development in this topic would advance these co-related research subjects.Owing to the wide application prospects and the high theoritical significance, recently, a great deal of research has been done on this topic. However, being a new research subject, existing methods are still far from satisfactory in practice. Through analyzing the requirement of practical applications and these existing works, we summarize the key issues of this topic as follows: (1) feature representation of human actions; (2) human action model; (3) computational complexity and (4) classification methods of human actions. Towards tackling these key issues, we propose new methods and insights for the feature learning, action modeling and classification algorithm of action recognition. The main contributions of this thesis are listed as follows:1. We propose an efficient and effecitive feature representation of actions and an action recognition method corresponding to this representation. Discriminative features are learnt from the dense simple feature descriptors of action video frames by finding the discriminative and statistically significant parts. Hence, the learnt features preserve the simple features'computational efficiency whereas they are more robust than the simple features. The feature learning approach can be easily extended to other applications, such as unsupervised anomaly detection and sequential action detection.2. We propose a hierarchical action grammar model. The model is characterized by: (i) it is based on discriminative high-order statistics of local features (e.g. spatiotemporal interest points). The statistics are learnt from the local spatiotemporal distribution of the local features which is capable of discerning different types of actions. (ii) The learned high-order statistics are represented as grammar rules, and a unified grammar model is proposed to depict the variations of actions using these grammar rules. (iii) A discriminative model is learnt from the generative action grammar model. In this manner, our model leverages the grammatical representation's expressive power and the discriminative model's potent discriminativity and computational efficiency.3. We propose a novel classification methodology, namely, Emerging Pattern Random Forest (EPRF). The method is built on the random forest classifier. By analyzing the statistical properties exhibited in the outputs of weak classifiers (random trees) of the random forest classifier, we learn a set of discriminative rules. Each rule combines the outputs of a subset of weak classifiers to provide a vote for the classification descision of the EPRF. The method takes into consideration the classification capability variations of each weak classifier in different data classes, and improve the performance of weak classifiers by combining them in the discriminative rules. Experimental results confirms that the proposed classifier outperforms the original random forest classifier.In summary, the thesis proposes new methods and insights to tackle the key issues in human action recognition research and enhance the applicability of the action recognition methods. In the experiments, we show that our methods improve the performance of existing methods, and have wide application prospects.