To reduce quantization error in vector quantization of Bag of Words (BoW) for video semantic detection and extract feature automatically and effectively, a new video semantic detection method based on Topographic Independent Component Analysis (TICA) and Gaussian Mixture Model (GMM) was proposed. Firstly, features of each video clip were extracted by TICA algorithm to learn complex invariant features from video clips. Secondly, the feature distribution of each video clip was described by GMM. Finally, a GMM supervector was created from GMM parameters and the GMM supervector for each shot was used as the input of an Support Vector Machine (SVM) for video semantic detection. A GMM can be regard as an extension of the BoW to a probabilistic framework, and thus, has less quantization error, better retaining the information in the original feature vectors. The experiments were conducted on the TRECVID 2012 and OV datasets. The experimental results show that compared with BoW and SIFT (Scale Invariant Feature Transform)-GMM algorithm, the proposed method can improve the mean average precision on both of the TRECVID 2012 and OV datasets for video semantic detection.