230 likes | 605 Views
Person De-Identification in Videos. Prachi Agrawal and P. J. Narayanan IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. 21, NO. 3, MARCH 2011. Person De-Identification in Videos. Outline. Introduction De-Identification: General Framework
E N D
Person De-Identification in Videos PrachiAgrawal and P. J. Narayanan IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. 21, NO. 3, MARCH 2011
Outline • Introduction • De-Identification: General Framework • De-Identification: Proposed Approach • Detect and Track • Segmentation • De-Identification • Experimental Results • Conclusions
Introduction • This has raised new concerns regarding the privacy of individuals. • Videos over the internet invaded our privacy. • Some technologies like Google street view, or surveillance video. • It is needed to person de-identification in videos.
Introduction • Face recognition and human detection are accurate recently. • But just black out the face or human will lose many information in the video. • In this paper, they use two blur methods to remain more information.
Different Scenarios and De-Identification • Casual videos • Public surveillance videos • Private surveillance videos
Subverting De-Identification • Reversing the de-identification transformation is the most obvious line of attack. • estimate the blurring function from the de-identified frames • Randomize is needed.
Storage of Videos • The safest approach is to de-identify the video at the capture-camera. • Some situation need the original video. • Final approach is to store the original video, with sufficiently hard encryption, along with the de-identified video.
Detect and Track • HOG based human detector. • patch-based recognition approach for object tracking by voting. • apply the human detector every F frames. • set F to 40 for our experiments
Segmentation • Multiple video tubes are formed if there are multiple people in the video. • voxels of size (x × y × t) in the spatial (x, y) and temporal (t) domains. (4*4*2)
Segmentation • U is the data term and V1, V2 are the smoothness terms corresponding to the intra-frame and inter-frame • The Gaussian mixture models (GMMs) are used for adequately modeling data points in the color space
Segmentation • The representative color vn for a voxelshould be chosen carefully. • distance D0 and D1 to the background and foreground • The pixels are sorted on the ratio D0/ D1 in the decreasing order. • D1 is low in mth pixel, seed foreground • D0 is low in (N-m)th pixel , seed background
De-Identification • exponential blur • Weight
De-Identification • line integral convolution (LIC)
Experimental Results • 97.2% and 7.8% hit rates in the case of person detector and face detector
Conclusion • presented a basic system to protect privacy against algorithmic and human recognition • We also conducted a user study to evaluate the effectiveness of our system. • characteristics are difficult to hide if familiarity is high to the user