Electronic image stabilization using multiple visual cues

Fast Electronic Digital Image Stabilization for Off-Road Navigation

“…The horizon is a very strong visual cue, present in almost every off-road situation. Yao et al [12] and Duric and Rosenfeld [4] used similar visual cues. The first step in feature extraction is to smooth the input image by convolving it with a 5 ϫ 5 Gaussian kernel operator.…”

“…Not compensating for the scale factor produces a similar effect; the smoothing techniques suggested by Duric and Rosenfeld [4] and Yao et al [12] could be applied to the remaining parameters. The advantage of such a scheme is that the current system could still be used without much change in performance, but a better scheme would use the tracked feature pairs to estimate the 3D rotation directly, but using iterative methods like those in Duric and Rosenfeld [4] and Irani et al [6] slows down the process significantly.…”

“…The latter approach has been implemented in our current system. Other schemes presented in [4,12] could be used to smooth some or all of the parameters.…”

“…Stabilization in 3D is achieved by de-rotating the frames, generating a translation-only sequence, or a sequence containing translation and low-frequency rotation (smoothed rotation). Yao et al [12] compensates for 3D rotation by tracking multiple visual cues, like distant points and horizon lines, using an extended Kalman filter for the estimation of the 3D motion parameters of interest. Both kinematic and kinetic models suitable for determining the smooth and oscillatory rotational motion components are considered, so that smoothed rotation can be also obtained.…”

Efficient Video Rectification and Stabilisation for Cell-Phones

“…Such approaches only compensate for the 3D rotational motion component, and neglect the translations, in a similar manner to mechanical stabilisation rigs (Yao et al, 1995). Rotational models estimate a compensatory rotational homography, either using instantaneous angular velocity (Yao et al, 1995) (differential form), or using inter-frame rotations, e.g. represented as unit quaternions (Morimoto and Chellappa, 1997).…”