Multiscale algorithm for reconstructing videos from streaming compressive measurements

Abstract

We propose a multiscale, iterative algorithm for reconstructing video signals from streaming compressive measurements. Our algorithm is based on the observation that, at the imaging sensor, many videos should have limited temporal bandwidth due to the spatial lowpass filtering that is inherent in typical imaging systems. Under modest assumptions about the motion of objects in the scene, this spatial filtering prevents the temporal complexity of the video from being arbitrarily high. Thus, even though streaming measurement systems may measure a video thousands of times per second, we propose an algorithm that only involves reconstructing a much lower rate stream of “anchor frames.” Our analysis of the temporal complexity of videos reveals an interesting tradeoff between the spatial resolution of the camera, the speed of any moving objects, and the temporal bandwidth of the video. We leverage this tradeoff in proposing a multiscale reconstruction algorithm that alternates between video reconstruction and motion estimation as it produces finer resolution estimates of the video.