Collective motion pattern inference via Locally Consistent Latent Dirichlet Allocation

Extracting motion descriptors in crowd videos is highly challenging due to scene clutter and serious occlusions. In this paper, Locally Consistent Latent Dirichlet Allocation (LC-LDA) model is proposed to learn collective motion patterns using tracklets and bag-of-words as low level features. The LC-LDA model implements a graph Laplacian operator to impose neighboring constraints to tracklets on a local manifold, which enforces the spatial–temporal coherence of tracklets in a high dimensional bag-of-word feature space. With initialization of clustering on a manifold, LC-LDA model improves the unsupervised inference capability and compactness of learned collective motion patterns. Experimental results on three public datasets indicate that LC-LDA based motion patterns can improve the trajectory clustering performance effectively.