Gender Classification Using Video Sequences of Body Sway Recorded by Overhead Camera

We extract time-series signals of local upper-body movements from overhead video of a standing person. Using spatio-temporal features of body sway, we investigate gender classification when larger movements such as walking are absent.

Some methods classify gender from the way people walk, but those methods are less suitable for people who are standing still. We focus on body sway: the small movements that occur naturally even while a person is standing. We investigate whether overhead video of these movements contains useful information for gender classification.

We divide the upper body into small regions and measure how each region moves over time. We then use frequency analysis to examine the slower and faster components of the sway. Combining information about where the movement occurs and how quickly it changes gives the model a way to describe body sway.

In experiments using recordings of 30 female and 30 male participants, the method achieved 90.3% gender-classification accuracy. The improvement over a method using measures based on medical data was statistically significant: it was unlikely to be explained by random variation alone. These results describe performance for the participants and conditions examined in this experiment.

• 神谷 卓也, 山口 優太, 中谷 真太朗, 西山 正志, 岩井 儀雄,
  頭上カメラから観測された身体動揺を用いた性別認識の精度評価,
  映像情報メディア学会誌, Vol. 75, No. 3, pp. 438 - 447, May 2021.
• Takuya Kamitani, Yuta Yamaguchi, Shintaro Nakatani, Masashi Nishiyama, Yoshio Iwai,
  Gender Classification Using Video Sequences of Body Sway Recorded by Overhead Camera,
  Proceedings of 25th International Conference on Pattern Recognition (ICPR), pp. 9196 - 9202, January 2021.

• Gender recognition using body sway

• Masaya Kojima, Masashi Nishiyama, Yoshio Iwai,
  Effect of Calf Muscle Fatigue on Person Identification Using Video Sequences of Body Sway,
  Proceedings of 20th International Symposium on Communications and Information Technologies (ISCIT), pp. 74 - 78, October 2021. [Best Student Paper Award]