De-genderization by body contours reshaping

This paper deals with privacy protection in video surveillance. More specifically, the main goal of this work is to make the gender of people no more recognizable while preserving enough information concerning body shape and motion of people for action classification.

We denote this processing as de-genderization. Regarding the current state-of-art methods, most of them have privacy filters only dedicated to de-identify people. These methods do not automatically imply the suppression of visual semantic traits such as gender. Therefore, we propose two approaches that modify the visual appearance of the body shape in order to de-genderize people while keeping the possibility to interpret the video. In both methods we start by extracting the contour points attached to the body shape of people in videos. Then we either mix the coordinates of the body shape and a predefined model, or we smooth the body shape by successive polygonal approximations based on convexity. Our results demonstrate that both proposed approaches protect the gender information while preserving the global body movement. The second approach based on convexity better preserves the visibility of human activities.