抄録

◎ジャ ハオ・櫻井和之・今岡　仁（NEC）

This paper presents a video face tracking method using face direction information. Since the previous methods use only past location information for prediction, the predicted location is often incorrect in scenes with occlusion or face direction change. In this paper, we solve this problem by combining two prediction models: one model is our face direction-based face tracking, which uses both current and past face direction information for prediction; another model is based on the conventional method using face location. As a result, mate separate rate, which indicates the ratio that a ground truth trajectory is interrupted, is decreased from 9.7% to 2.3% for the outdoor surveillance dataset. The proposed method is suitable for more general environments.