Abstract: For the characteristics of video feature diversity and sparse dictionary redundancy,a method of video semantics based on discriminative features-blocked sparse representation is presented in this paper.Firstly,various features are extracted from video segments as actual requirement.Then each blocked dictionary for the sparse representation is constructed in terms of the dimensional size of each feature.The dictionaries are optimized by embedding discriminative criterion into K-SVD algorithm in order to improve category identification ability of sparse representation feature.In the semantic analysis for video segment,the sparse features of any kinds of primitive features are obtained by the optimized dictionaries.The weighted KNN is adopted for classification analysis of these sparse features.Finally,the video segment semantic is analyzed by fusing analysis of the sparse features according to the decision support degrees.The experimental results show that the proposed method can effectively improve the accuracy and speed of video semantic analysis.