Human Action Recognition Based On Spatio-temporal Interest Points

Recognition of human actions has already been an active area in the computer vision domainand techniques related have been applied in plenty of fields such as smart surveillance, motionanalysis and virtual reality. The dissertation proposes a kind of method to recognize the single andtwo-people interaction activity on the foundation of the detected spatio-temporal interest pointsfrom the video sequences.Firstly, this dissertation describes the general three-dimensional spatial-temporal interest pointsdetectors and descriptors. Due to the weakness of frequently-used interest points detect method, thedissertation proposes a new detected method based of2D Gabor filter. Next, they briefly introducethe principle of clustering algorithm and the bag-of-words models and their application in actionrecognition field.Secondly, they employ the bag-of-words model to convert the spatio-temporal feature intohistogram feature and employ the probabilistic theme model, specifically, PLSA and LDA algorithmto analyze the topic distribution of the video sequences. They propose the labeled LDA (L-LDA),which is an extension of Latent Dirichelet Allocation model. The L-LDA adds a label layer on thebasis of LDA to label the category of the train video sequences. L-LDA assigns the latent topicvariable in the model to the specific action categorization automatically, so it can help to estimatethe model parameters more reasonably, accurately and fast.At last, they employ Markov Logic Network to analyze and recognize two-people interactionactivity. Markov Logic Network combines the ability of probabilistic graphical model to infer theprobability flexibly with the ability of statistical relational model to deal with indeterminacy. Theytreat two-people interaction activity as two single person actions separately. Combined the singleaction class with the semantic information, at last they employ the Markov Logic Network to infertwo-people interaction activity.The experiment of recognizing single people action and two-people interaction activity isseparately based on KTH database and UT-interaction database, and the results show satisfyingrecognition rates are obtained.