Dynamically visual learning for people identification with sparsely distributed cameras

We propose a dynamic visual learning method that aims to identify people by using sparsely distributed multiple surveillance cameras. In the proposed method, virtual viewpoint images are synthesized by interpolating the sparsely distributed images with a simple 3D shape model of the human head, so that virtual densely distributed multiple images can be obtained. The multiple images generate an initial eigenspace in the initial learning step. In the following additional learning step, other distributed cameras capture additional images that update the eigenspace to improve the recognition performance. The discernment capability for personal identification of the proposed method is demonstrated experimentally.