Fisheye camera modeling for human segmentation refinement in indoor videos

In this paper, we concentrate on refining the results of segmenting human presence from indoors videos acquired by a fisheye camera, using a 3D mathematical model of the camera. The model has been calibrated according to the specific indoor environment that is being monitored. Human segmentation is implemented using a standard established technique. The fisheye camera used for video acquisition is modeled using a spherical element, while the parameters of the camera model are determined only once, using the correspondence of a number of user-defined landmarks, both in real world coordinates and on the acquired video frame. Subsequently, each pixel of the video frame is inversely mapped to the direction of view in the real world and the relevant data are stored in look-up tables for very fast utilization in real-time video processing. The proposed fisheye camera model enables the inference of possible real world positions of a segmented cluster of pixels in the video frame. In this work, we utilize the constructed camera model to achieve a simple geometric reasoning that corrects gaps and mistakes of the human figure segmentation. Initial results are also presented for a small number of video sequences, which prove the efficiency of the proposed method.