[3Xin4] Poster session 1

[3Xin4-15] Camera rotation estimation using person region in image with less orthogonal components

〇Takahiro Yoshioka¹, Jinta Sakai², Takeshi Konno¹ (1.FUJITSU LIMITED, 2.Univ. of Tsukuba)

Keywords:Camera calibration, Human height estimation

Security solutions aim to detect, track, and identify people using images from multiple cameras. This study focused on human height information to accurately search for a person. Since the height information is basically constant regardless of the installation of the camera or shooting date, it becomes an important clue. In our previous study, we proposed a method to estimate human height by estimating camera parameters using pose of a person under a limited condition ignoring roll rotation of the camera. When there is roll rotation in the image, it affects the error of estimated height. Therefore, we propose a method to estimate rotation by using orthogonal components including human area. After we applied the proposed method and evaluate accuracies of human height estimation. As a result, we confirmed that the error of human height improved about 40% at the largest. In the future, the human height estimation result will be applied to the person search, and we aim to contribute to reducing the workload of investigators.