Research On Probabilistic Graphical Model Based Human Action Recognition Algorithm

Visual based human action recognition has already become one of the most activeresearch areas in computer vision. It has a wide spectrum of promising applications in manyareas such as smart surveillance, human-computer interaction, sports performance analysisand video content retrieval. Recognizing human actions correctly from arbitrarily cameraviewpoint is a meaningful task. However, it is very difficult due to the challenges, such as thelimited robustness of action representation, the complex parameter space during actionmodeling and the lacking view-space information of action models.According to the above three challenging problems, the researches on view-invariantaction recognition are discussed in this dissertation.Firstly, a mixed feature representation by combining the bag of interest point words inshot length-based video and the amplitude histogram of optical flow is proposed based on thetraditional bag of interest point words. This feature representation is robust to the minorvariations of viewpoint and easy to implement. At the same time, this representation containsrich human motion information and anti-noise-interference ability.Secondly, in order to address the issue of the complex parameter space during actionmodeling, a multi-view space probability combination Hidden Markov Models algorithmbased on view space partitioning is proposed and utilized for view-invariant actionrecognition. This algorithm effectively reduces the complexity of parameter space andincreases the accuracy of the action recognition by training human action models in eachsub-view space and recognizing action via the probability weighted combination.Finally, in order to solve the challenge of the lacking view-space information in actionmodels, on the basis of multi-view space Hidden Markov Models, a novel Hidden MarkovModels approach with multi-view space state transition is proposed. This approach couldstrengthen the viewpoint constraints among the human action models in the sub-view spaceand improve the recognition performance for viewpoint changing problem by introducing the multi-view space state transition probability.