Research On Recognition Of Human Actions For Interior Scenes

Human Action Recognition Technology is used to analyze video-sequences and detect human action automatically. The nature of action recognition is a time-varying data classification. This paper analyzes characteristics of human behaviors in daily life. Based on the comparison of exist recognition algorithms, Achievements of interior scenes feature extraction and action learning algorithms are as follows:(1)Feature extraction:There are two problems exits, few local temporally spatial features and weakness among mid-level feature expression. To solve them, this thesis combines the depth information in videos and presents a Spatial Temporal Depth Feature (STDF).This feature could provide more discriminated information in intense areas. The principle of STDF is depth information to detect these kinetic regions and then calculate optical flows to be energy function. STDF samples interesting points on kinetic region in Gaussian distribution through this function. After feature extraction, the practical experiment uses SVM classifier to distinguish actions. At last, the result shows that the average accuracy rate on human action recognition in SwustDepth data set can reach 92%.(2)Unsupervised Learning of continuous action classification:There is a shortage on current algorithms that they could only recognize video pieces already been cut into individual actions. This paper proposes an unsupervised action recognition algorithm to combine spectral clustering and topic model. It uses spectral clustering to generate visual words, which avoid local extreme value and sensitive originals. This algorithm discovers the categories in statistical text literature:pLSA and LDA model. As a result, this algorithm could recognize and locate human actions in a long video containing multiple simple actions.