Human Action Recognition in Videos Using Intermediate Matching Kernel

Human action recognition can be considered as the process of labelling the videos with the corre-\ud sponding action labels. Coming to the elds of computer vision, video sensing this has become an\ud important area of research. There are a lot of factors such as recording environment,intra class and\ud inter class variations,realistic action ambiguities and varying length of actions in the videos which\ud make this problem more challenging\ud Videos containing human actions can be considered as the varying length patterns because the\ud actions in videos may last for dierent duration. In this thesis the issue of varying length patterns\ud is being addressed. To solve this issue a paradigm of building intermediate matching kernel as a\ud dynamic is used so that the similarity among the patterns of varying length can be obtained. The\ud idea of the intermediate matching kernel is using a generative model as a reference and obtain\ud the similarity between the videos. A video is a sequence of frames which can be represented as\ud a sequence of feature vectors and so hidden markov model is used as the generative model as it\ud captures the stochastic information. The complete idea of this thesis can be described as building\ud intermediate matching kernels using hidden markov model as generative model over which the SVM\ud is used as a descriminative model for calssifying the actions based on the computed kernels. This\ud idea is evaluated on the standard datasets like KTH, UCF50 and HMDB5