Video-based Human Action Recognition

Analysis of human activities has always remained a topic of great interest in computer vision, and humanaction recognition is a challenging research topic in this field．Human action recognition is critical for a widerange of applications, such as video surveillance, man-machine interface, digital entertainment, intelligentinterface, environmental control and monitoring, sport video annotation, etc. The main contributions are asfollows：This thesis aims to extract a good feature descriptor, which not only describes motion information but alsocontains the local information of strcture and appearance. The histograms of oriented optical flow (HOOF)describe motion features of certain frames in video and the histogram of oriented gradient (HOG) can be used tocatch the local information of structure and appearance. In this thesis, these two features are weighted fusion to geta good festure descriptor, which is histogram of weighted optical flow (HOWOF). This feature descriptor containsnot only the motion information of human action, as well as the silhouette of the human figure, which is a goodfeature descriptor and has laid a good foundation for behavior classification and recognition of human actions inthe follow-up process.System theoretic approaches to action recognition model the dynamics of a scene with linear dynamicalsystems (LDSs) and perform classification using metrics on the space of LDSs, e.g. Binet-Cauchy kernels.However, such approaches are only applicable to time series data living in a Euclidean space, e.g. joint trajectoriesextracted from motion capture data or feature point trajectories extracted from video. Much of the success ofrecent object recognition techniques relies on the use of more complex feature descriptors, such as SIFTdescriptors or HOG descriptors, which are essentially histograms. Since histograms live in a non-Euclidean space,we can no longer model their temporal evolution with LDSs, nor can we classify them using a metric for LDSs.Therefore, this thesis proposes a generalization of the Binet-Cauchy kernels to nonlinear dynamical systems(NLDS) whose output lives in a non-Euclidean space, e.g. the space of histograms, and with the method of kerneldiatance measurement for classifying videos.