A Laplacian structured representation model in subspace clustering for enhanced motion segmentation

Segmenting a moving object from its background is a fundamental step in many computer vision applications ranging from visual surveillance to multimedia image analysis. Although subspace clustering based motion segmentation methods (such as Sparse Subspace Clustering (SSC) and Low-Rank Representation (LRR)) have achieved state-of-the-art performance, they have suffered from the loss of local structure problem, i.e., local similar features may be encoded as totally different codes due to the overcomplete codebooks. Such instability may harm the connectivity of the similarity graph and affect the performance of clustering algorithms finally. To remedy this issue, we propose a Laplacian structured representation model to enhance the representation-based clustering methods by importing local feature similarity prior information to guide the encoding process, and then develop an efficient Alternating Direction Method of Multipliers (ADMM) algorithm for optimization. Two improved subspace clustering methods, the Enhanced Sparse Subspace Clustering (E-SSC) and Enhanced Low Rank Representation (E-LRR), are devised in this work. Experiments on Hopkins 155 motion segmentation dataset and airport dataset demonstrate the advantage of our proposed model over state-of-the-art methods, and achieve 0.77% and 0.85% segmentation error rate, respectively.