Superpixel Segmentation of Outdoor Webcams to Infer Scene Structure

Abstract

Understanding an outdoor scene’s 3-D structure has applications in several fields, including surveillance and computer graphics. Scene elements’ time-series brightness gives insight to their geometric orientation; and thus the 3-D structure of the overall scene. Previous works have studied the time-series brightness of individual pixels. However, there are limitations with this approach. Pixels are often quite noisy, and can require a lot of memory. This thesis explores the use of superpixels to address these issues. Superpixels, an approach to image segmentation, over-segment a scene but attempt to ensure that each segment lies on only one scene element. Applying superpixels to webcams reduces the effect of noise on pixels’ time-series brightness, and conserves memory by reducing the number of pixel “entities”. This thesis explores methods of solving for a superpixel’s surface normal, and demonstrates that the time at which maximum brightness is achieved serves as a basic indicator of geographic orientation.