A Direct Method to Self-Calibrate a Surveillance Camera by Observing a Walking Pedestrian

Recent efforts show that it is possible to calibrate a surveillance camera simply from observing a walking human. This procedure can be seen as a special application of the camera self-calibration technique. Several methods have been proposed along this line, but most of them have certain restrictions, such as require the human walking at a constant speed, or require two orthogonal lines marked on the ground, etc. This has hindered their applicability. In this paper we propose a new method that removes most of these restrictions. By clever uses of the cross-ratio relationship in projective geometry, our method shows it is possible to directly estimate a full 3x4 camera projection matrix without first decomposing it into physical parameters like focal-length, optical center, etc. Extensive experiments on real data show our algorithm performs well in real situations.