Multi-Camera Joint Self-Calibration from Observations of Pedestrians

Liuzhen Su, Ye Lin, Lihua Zhang

Abstract: Calculating camera parameters by vanishing points is a widely used self-calibration method, but this method is usually accompanied by large geometric errors. In this paper, we propose a novel multi-camera joint self-calibration method based on vanishing points estimation and evolutionary algorithm, which only uses a few human walking videos as input. The vanishing point and vanishing line are clustered and regressed respectively according to the torso detection results instead of head and foot points, so as to reduce the influence of human posture changes. Then the initial camera parameters of each camera calculated by vanishing points are simultaneously optimized through the joint geometric error. Experiments on the open data set show the accuracy and flexibility of the proposed method.