| Reliable motion estimation in videos is of crucial importance for background iden- tification, object tracking, action recognition, event analysis, self-navigation, etc. Re- constructing the motion field in the 2D image plane is very challenging, due to variations in image quality, scene geometry, lighting condition, and most importantly, camera jit- tering. Traditional optical flow models assume consistent image brightness and smooth motion field, which are violated by unstable illumination and motion discontinuities that are common in real world videos.;To recognize observer (or camera) motion robustly in complex, realistic scenarios, we propose a biologically-inspired motion estimation system to overcome issues posed by real world videos. The bottom-up model is inspired from the infrastructure as well as functionalities of human dorsal pathway, and the hierarchical processing stream can be divided into three stages: 1) spatio-temporal processing for local motion, 2) recogni- tion for global motion patterns (camera motion), and 3) preemptive estimation of object motion. To extract effective and meaningful motion features, we apply a series of steer- able, spatio-temporal filters to detect local motion at different speeds and directions, in a way that's selective of motion velocity. The intermediate response maps are cal- ibrated and combined to estimate dense motion fields in local regions, and then, local motions along two orthogonal axes are aggregated for recognizing planar, radial and circular patterns of global motion. We evaluate the model with an extensive, realistic video database that collected by hand with a mobile device (iPad) and the video content varies in scene geometry, lighting condition, view perspective and depth. We achieved high quality result and demonstrated that this bottom-up model is capable of extracting high-level semantic knowledge regarding self motion in realistic scenes.;Once the global motion is known, we segment objects from moving backgrounds by compensating for camera motion. For videos captured with non-stationary cam- eras, we consider global motion as a combination of camera motion (background) and object motion (foreground). To estimate foreground motion, we exploit corollary dis- charge mechanism of biological systems and estimate motion preemptively. Since back- ground motions for each pixel are collectively introduced by camera movements, we apply spatial-temporal averaging to estimate the background motion at pixel level, and the initial estimation of foreground motion is derived by comparing global motion and background motion at multiple spatial levels. The real frame signals are compared with those derived by forward predictions, refining estimations for object motion. This mo- tion detection system is applied to detect objects with cluttered, moving backgrounds and is proved to be efficient in locating independently moving, non-rigid regions.;The core contribution of this thesis is the invention of a robust motion estimation system for complicated real world videos, with challenges by real sensor noise, complex natural scenes, variations in illumination and depth, and motion discontinuities. The overall system demonstrates biological plausibility and holds great potential for other applications, such as camera motion removal, heading estimation, obstacle avoidance, route planning, and vision-based navigational assistance, etc. |
