Achieving $\epsilon$-Object Indistinguishability in Surveillance Videos Through Trajectory Randomization

Abstract

Data has become an integral part of our digital lives. Specifically, there is an explosive growth in the application of video data. Video information is quite different from other data formats in the sense that it possesses unique characteristics such as high dimensions, complex content, and multiple forms of representation. All these properties make the protection of privacy in videos a complex and challenging task. Simple obscuration techniques cannot address the conclusions or deductions extrapolated from the background information of the entities in the video. In this work, we explore a video sanitization technique that generates synthetic videos following the perturbation of the objects of interest. In our model, we combine the naive detect and obscure technique with randomization in the presence of the objects of interest in each frame and their trajectories. Essentially, our holistic model fulfills the privacy notion of $\epsilon$-object indistinguishability. The generated videos achieve our aim of preserving privacy while being accurate enough for utility analysis. We tested our system on the MOT16 videos dataset and observed a reasonable count of 20% lost objects, mean square error ranging in $[0.2-0.3]$, and trajectories deviation between $[0.2-0.6]$.