Key Points
We combine face images of the same person captured by multiple cameras to form a distribution for each individual. Matching these distributions, rather than individual images, helps handle changes in facial pose and illumination as people move.
Abstract
Recognizing faces in a moving crowd is difficult. Other people may hide a face, and its direction and lighting can change. This study proposes using several cameras to capture near-frontal views and judging a person's face from a collection of images.
Instead of matching one image at a time, we combine images of the same person from different cameras. Together, they represent the range of ways that person's face can appear. We compare a newly captured collection with the collections of people registered in advance. This helps recognize the same person despite changes in pose or illumination. The method is called Hierarchical Image-Set Matching.
Tests using videos containing 349 people showed better recognition performance than methods that matched each frame separately or used an image collection from only one camera. The results demonstrate the benefit of combining multiple views and images when recognizing people on the move.
Publication
- Masashi Nishiyama, Mayumi Yuasa, Tomoyuki Shibata,
Tomokazu Wakasugi, Tomokazu Kawahara, and Osamu Yamaguchi,
Recognizing Faces of Moving People by Hierarchical Image-Set Matching, Proceedings of IEEE Computer Society Workshop on Biometrics, pp. 1 - 8, June 2007. [Best Paper Award]
Publications (Japanese)
- 西山 正志, 湯浅 真由美, 柴田 智行, 若杉 智和, 山口 修,
顔画像の階層的な対応付けを用いた複数歩行者の認識,
電子情報通信学会論文誌 D, Vol. J90-D, No. 8, pp. 2191 - 2201, 2007. - 西山 正志, 湯浅 真由美, 若杉 智和, 柴田 智行, 山口 修,
断片的な動画像の対応付けを利用した歩行者認識,
第12回 画像センシングシンポジウム (SSII), pp. 231 - 238, June 2006. [優秀論文賞]
