Real-time Image Stabilizer

PROJECT DESCRIPTION:

Objective

The goal of this study is to design a zero-phase adaptive filter in order to accurately separate voluntary and involuntary camera movements in real-time. The filtering of the erroneous motion allows to compensate for it while preserving the user's intentional movements.

Abstract

Handholding a digital video camera or filming a video from a moving vehicle most often produces shaky videos. Nowadays, digital video cameras are often equipped with a built-in image stabilizing system, which can greatly reduce or virtually eliminate the vibration caused by camera shake. Because a user may intentionally move the camera, a good image stabilizing system must be able to distinguish intended motion from undesired motion.

The goal of this study is to design a zero-phase adaptive filter in order to accurately separate voluntary and involuntary camera movements in real-time. The filtering of the erroneous motion allows to compensate for it while preserving the user's intentional movements. We demonstrate our method in one dimension using a shaker to generate a vertical undesired motion. We test our method with a vertical video image stabilization system, which reduces the RMSE of the erroneous motion with respect to the ground truth by 75%.

Current Status

In this work, we investigated image stabilizing systems for video cameras. In particular, we proposed a method which enables such systems to distinguish intended motion from undesired motion. Our method allows these systems to accurately compensate for the undesired camera movements while preserving the users intentional movements. To achieve this, we proposed the adaptive phase-balancing bandpass IIR filter. The presented method uses the WFLC algorithm to adapt to the changing frequency of the erroneous motion and has the benefit of filtering it in real-time without introducing any delay. We tested the performance of our filter by conducting an experiment in which we simulated the undesired camera movements that would result from filming from a moving car. For this purpose, we designed a stand to attach a motion sensor to a video camera and implemented a one-dimensional video image stabilization system.

From our experiment, we observed that our adaptive phase-balancing bandpass IIR filter achieves zero-phase filtering of the erroneous motion. Furthermore, our video image stabilization system can successfully compensate for the undesired camera movements, reducing the RMSE of the erroneous motion with respect to the ground truth by 75%. Our video image stabilization system was unable to produce a perfectly steady video because we did not calibrate the motion sensor as would normally be required. This issue was partially resolved by extracting a ground truth from the raw video. This information was used to correct the amplitude and phase of the displacement measurements. Moreover, our method for stabilizing video images suffers from rounding errors.

Calibrating the motion sensor would definitely improve the accuracy of our video image stabilization system. In addition, our filter should be incorporated within a video camera and tested with a more precise image stabilizing system such as an OIS system. The resulting hardware implementation would also allow us to test the performance of the filter in real-time. Because of its adaptive nature, our filter can be used in a wide range of systems. For example, it could be used for active real-time canceling of pathological or physiological tremor. In fact, our method can be used in any application where zero-phase filtering of quasiperiodic signals is crucial.

GRANT:

• –

PERSONNEL:

Dr. ANG Wei Tech
Ph.D.

Assistant Professor

wtang@ntu.edu.sg