Multimodal human detection by sparse feature pursuit

Human detection from multimodal image is a challenging task of information extraction and plays a striking important role for the later steps such as classification, recognition, tracking, and so forth. This paper describes an innovative sparse feature-based approach for human detection using the multimodal image. Firstly we consider a human as sparse feature which moves with multimodal image sequences. And afterwards the problem of moving human estimation can be formulated as decomposition of a matrix into a sparse human matrix and a low-rank background matrix. Furthermore, both of the components are exactly recovered by solving convex optimization problem. Finally the sparse feature that contains human is reconstructed to generate the human map. Experimental results on the real multimodal image from a novel 2D/3D vision system verify the effectiveness of our proposed method. Meanwhile the results yield the potential application of matrix decomposition for various multimodal data analysis.