Signed coded exposure sequences for velocity and shape estimation from a single photo

Abstract

In this thesis, I analyze the benefits of signed coded exposure for velocity and shape detection of moving objects. It has been shown that coded exposures enhance deblurring of motion blurred photos [28]. However, these non-negative binary codes (1 or 0) only suggest opening and closing of the shutter to allow or prevent light from entering the camera. Signed codes (+1 or -1) for camera exposures offer accumulation or removal of light over the course of a single exposure. I show that signed codes provide dramatic benefits over unsigned code for motion estimation due to better frequency domain properties and auto-correlation characteristics. I analyze the space of such codes with invertibility analysis and a cross-correlation metric. Motion estimation is important to a number of computer vision problems such as tracking, segmentation and recognition. New emerging hardware, in the commercial and research domains, provides signed coded for exposure, but their full capabilities have not been explored. Part of my efforts involved experimenting with the electronics of such cameras. The emphasis in this thesis is on the computational aspects of a framework which employs new codes for motion parameter estimation. I demonstrate the ideas on a variety of synthetic images and real-world photographs. I hope the cameras and theory of signed coded exposure will be a new motion-analysis tool in the field of computational photography.