Automated 3D object modeling from aerial video imagery

Research in physically accurate 3D modeling of a scene is gaining momentum because of its far reaching applications in civilian and defense sectors. The modeled 3D scene must conform both geometrically and spectrally to the real world for all the applications. Geometric modeling of a scene can be achieved in many ways of which the two most popular methods are a) using multiple 2D passive images of the scene also called as stereo vision and b) using 3D point clouds like Lidar (Light detection and ranging) data. In this research work, we derive the 3D models of objects in a scene using passive aerial video imagery. At present, this geometric modeling requires a lot of manual intervention due to a variety of factors like sensor noise, low contrast conditions during image capture, etc. Hence long time periods, in the order of weeks and months, are required to model even a small scene. This thesis focuses on automating the process of geometric modeling of objects in a scene from passive aerial video imagery. The aerial video frames are stitched into stereo mosaics. These stereo mosaics not only provide the elevation information of a scene but also act as good 3D visualization tools. The 3D information obtained from the stereo mosaics is used to identify the various 3D objects, especially man-made buildings using probabilistic inference provided by Bayesian Networks. The initial 3D building models are further optimized by projecting them on to the individual video frames. The limitations of the state-of-art technology in attaining these goals are presented along with the techniques to overcome them. The improvement that can be achieved in the accuracy of the 3D models when Lidar data is fused with aerial video during the object identification process is also examined.