|M.Sc Student||Levi Eliyahu|
|Subject||Image Based 3-D Modeling of complex dynamic scenes|
|Department||Department of Civil and Environmental Engineering||Supervisor||Professor Sagi Filin|
|Full Thesis text|
The application of terrestrial photogrammetry for documentation and reconstruction of sites is widely known, and is driven by the ease of data collection, availability of consumer grade cameras, the ability to avoid physical contact with mapped objects, and the existence of well-established photogrammetric documentation methodologies.
Although image-based dense 3-D modeling has been studied extensively in the past, including stereo-vision and multi-view approaches, they assume that the imaged scenes remain static throughout the data acquisition campaign. Nonetheless, a significant part of real-world scenes is dynamic and may contain more than a few independently moving object. To facilitate scene modeling in the presence of such complex environment, the objective of this research is to study image-based 3-D reconstruction of dynamic scenes.
Existing approaches to model such scenario can be roughly divided into two classes: one in which multiple statically-mounted cameras acquire the data and another in which a single hand-held moving camera is employed. All approaches require continuous video stream so that motion is limited, and when multiple cameras are employed, they are restricted to highly controlled conditions. In the proposed approach we overcome both restrictions by utilizing a keypoints-based motion modeling. As opposed to existing approaches the proposed methodology shows that motion can then be estimated even when the camera-object motion is relatively large. It is therefore unrestricted to neither mounted imaging-configurations nor continuous video streams. As extracted keypoints are sparse by nature and dense representation of the imaged scene cannot be established just based on them, we adopt an expansion methodology from a multi-view modeling approach and modify it to a dynamic environment. Therefore, a dense 3-D reconstruction model which handles motion within the scene is the outcome of this thesis.
The study shows the application of the proposed method on manmade scene with dynamic object in it. Results show that the application of the method provides an efficient and detailed 3-D model of both moving and static objects.