Structure from Motion

We propose to develop a superior 3D surface reconstruction algorithm that operates on monocular aerial video. High quality pose estimates from inertial data will be used to construct dense depth maps of the 3D environment. Each 3D pixel in the reconstruction will be extracted from multiple views in the redundant video stream to minimize the error in the reconstruction. We will refine and adapt the design of our reconstruction algorithms to make them suitable for hard-silicon logic implementation such as FPGAs or ASICs. The algorithms will be optimized for massively parallel image processing overseen by relatively few processors of greater sophistication. A two pass workflow will be implemented where terrain appearing in the field of view will first be reconstructed with relatively low depth resolution, followed by successively more accurate reconstruction as a region traverses the field and available baseline and image count on subject increase. This also allows resolution to be variably dependent on processing power, groundspeed, and field of view. We will construct high quality poses for demonstration of our Phase II technology using a SfM approach which fuses GPS, inertial data and sparse visual features.