Robust extrinsic calibration from pedestrians

Abstract

Calibration from pedestrians in the single-camera setup is studied in this paper. Previous relevant methods are mainly designed in simple scenes with a single pedestrian. The objective of this paper is to improve the calibration performance in complex scenes with crowds. Different methods in previous work are evaluated and compared. Under the assumption of known camera intrinsic parameters, a novel metric for vanishing line estimation is proposed, and WCS reference is incorporated into the calibration method. Experiments are conducted on the PETS′09 dataset with crowds and the Town Center dataset in real world. The experimental results prove that the proposed metric outperforms previous metrics in the calibration, and the refinements of WCS reference can further improve the calibration accuracy.