Electronic stabilization and feature tracking in long image sequences

This dissertation is concerned with processing of visual motion with application to off-road vehicular navigation. Several aspects of the problem are investigated. First, we consider the estimation of total rotation from a sequence, useful for image stabilization. This procedure is important for motion analysis, idependently moving object detection as well as the recovery of other structural information. We exploit the dynamic nature of a sequence and use multiple visual cues to perform image stabilization. Depending on the knowledge of intrinsic parameters such as the focal length and the field of view, both calibrated and uncalibrated stabilization schemes are designed. The residual motion in a stabilized sequence is also analyzed. Next we address the issue of selective stabilization, defined as the separation of the smooth rotation and the residual oscillatory rotation. In off-road vehicular navigation, in addition to smooth motion, a vehicle exhibits residual vibrations. These residual oscillatory components often affect the interpretation of visual information. We incorporate a kinetic model to explicitly account for the phenomena of vibration. A maneuver detection scheme, for detecting the beginning and end of smooth rotation, is designed to facilitate the selective stabilization. The structure parameters are consequently estimated in a less perturbed frame of reference. Finally, we study the problem of feature correspondence. Tracking feature points over a sequence has been a critical procedure in exploiting an image sequence. We propose a localized feature point tracking algorithm. The method employs a 2-D kinematic model and relies on a Probabilistic Data Association Filter for the estimation of inter-frame motion. Corresponding points are identified to sub-pixel accuracy and an Extended Kalman Filter is employed to process the new data. The ability to dynamically include new feature points from subsequent frames also makes the algorithm suitable for structure from motion and tracking over a sequence.