Occlusion-based computational periscopy with consumer cameras

Abstract

The ability to form images of scenes hidden from direct view would be advantageous in many applications – from improved motion planning and collision avoidance in autonomous navigation to enhanced danger anticipation for first-responders in search-and-rescue missions. Recent techniques for imaging around corners have mostly relied on time-of-flight measurements of light propagation, necessitating the use of expensive, specialized optical systems. In this work, we demonstrate how to form images of hidden scenes from intensity-only measurements of the light reaching a visible surface from the hidden scene. Our approach exploits the penumbra cast by an opaque occluding object onto a visible surface. Specifically, we present a physical model that relates the measured photograph to the radiosity of the hidden scene and the visibility function due to the opaque occluder. For a given scene–occluder setup, we characterize the parts of the hidden region for which the physical model is well-conditioned for inversion – i.e., the computational field of view (CFOV) of the imaging system. This concept of CFOV is further verified through the Cram´er–Rao bound of the hidden-scene estimation problem. Finally, we present a two-step computational method for recovering the occluder and the scene behind it. We demonstrate the effectiveness of the proposed method using both synthetic and experimentally measured data.