Title: Temporal ghost Fourier compressive inference camera

Abstract: The need for real-time processing fast moving objects in machine vision requires the cooperation of high frame rate camera and a large amount of computing resources. The cost, high detection bandwidth requirements, data and computation burden limit the wide applications of high frame rate machine vision. Compressive Video Sensing (CVS) allows capturing events at much higher frame rate with the slow camera, by reconstructing a frame sequence from a coded single image. At the same time, complex frame sequence reconstruction algorithms in CVS pose challenges for computing resources. Even though the reconstruction process is low computational complexity, image-dependent machine vision algorithms also suffers from a large amount of computing energy consumption. Here we present a novel CVS camera termed Temporal Ghost Fourier Compressive Inference Camera (TGIC), which provides a framework to minimize the data and computational burden simultaneously. The core of TGIC is co-design CVS encoding and machine vision algorithms both in optical domain. TGIC acquires pixel-wise temporal Fourier spectrum in one frame, and applies simple inverse fast Fourier transform algorithm to get the desired video. By implementing pre-designed optical Fourier sampling schemes, specific machine vision tasks can be accomplished in optical domain. In fact, the data captured by TGIC is the results of traditional machine vision algorithms derived from the video, therefore the computation resources will be greatly saved. In the experiments, we can recover 160 frames in 0.1s single exposure with 16x frame rate gain (periodic motion up to 2002 frames, 1000x frame rate gain), and typical machine vision applications such as moving object detection, tracking and extraction are also demonstrated.