Abstract:A hierarchical adaptive clustering based group detection algorithm is proposed for the crowded scenes involving multiple complex motion modalities.Gaussian Mixture Models based background subtraction algorithm and the adaptive initialization clustering algorithm are the key to the algorithm.Group detection is implemented by merging spatiotemporal features of salient points into different layers of the model.Our dataset is built by varieties of in-door and out-door real scene videos.The proposed algorithm outperforms many other algorithms in terms of its effectiveness,reliability and superiority by experimental comparisons.