Dynamic scene analysis based on the topic model
Yawen Fan, Shibao Zheng · 2013
In this paper, a framework based on the topic model is proposed for dynamic scene analysis. Firstly, low-level motion features are detected and denoised. The residual low level feature is then mapped into visual words using a novel adaptive quantization method. The first level latent Dirichlet allocation(LDA) model is applied to automatically cluster visual words into atomic activities. Afterwards, the second level latent Dirichlet allocation model is used to cluster atom activity into interactions. Therefore video clips are represented as a mixture of interactions. The results of the real world traffic datasets demonstrate the effectiveness of the proposed method.