Person de-identification in activity videos

Person identification based on gait recognition has been extensively studied in the last two decades, while information appearing in different action types (like bend) has been recently exploited to this end. However, in most application scenarios it is sufficient to recognize the performed activity, whereas the ID of persons performing activities is not important. Since the same human body representations, e.g., body silhouettes, can be employed for both tasks, there is a need to automatically create privacy preserving representations. We have applied 2D Gaussian filtering to obfuscate the human body silhouettes that implicate information about the person ID. We have experimentally showed how the use of filtering affects the person identification and action recognition performance in different camera setups formed by an arbitrary number of cameras. In addition, the discriminative ability of different activities is examined and discussed in order to detect cases in which it is possible to apply Gaussian filter with a greater variance.