Jeremy Ryan Schiff and Ken Goldberg
National Science Foundation 0535218, National Science Foundation 0424422 and National Institute of Health
Motivated by security concerns, an emerging class of digital video cameras provides unprecedented ability to zoom in and capture high-resolution video images within a wide field of view. This capability is desirable for many practical applications, but also raises significant privacy concerns. We utilize wearable "markers" that can be detected by image processing software in real time to provide privacy. The example images illustrate yellow construction hats for our markers.
To determine the marker's location in each image, we train the system on a set of 200 marker pixels, and a set of 200 non-marker pixels and project the sets into a nine-dimensional color space. We use a statistical learning and classification technique, AdaBoost , to generate a final classifier described as a linear function of weak hypotheses. We then cluster pixels according to connected components, and label all clusters within specified area thresholds as the marker. Finally, we overlay a colored dot at a location relative to the detected marker. Our system produces a new video stream with these overlaid dots, thereby preserving privacy. This is our preliminary approach; our next steps will include temporal modeling and adaptive feature tracking.
Figure 1: An example of video frame input to our system
Figure 2: An example of video frame output from our system, using construction hats as our markers
- Y. Freund and R. E. Schapire, "A Decision-Theoretic Generalization of On-line Learning and an Application to Boosting," Computational Learning Theory, Eurocolt 1995, Springer-Verlag, 1995, pp. 23-37.