FPGA accelerated video-compression for cloud-based vision sensors

Vision Sensors for Cloud Applications have to drastically reduce the amount of image data that is transferred and stored. Due to the fact that such sensors are mostly bulk products, they must be optimized according to their performance, price, size and energy consumption. We present an approach for video compression that fulfills these needs by using FPGAs and allows scalability with respect to compression factor and compression rate. The underlying new algorithm is designed to be generic for (a) compressing a sequence of images in the JPEG standard and (b) extending the base system to a full video encoder using reference frames and prediction frames. The main idea is not to calculate the differences between consecutive video-frames directly, but between their spectra to eliminate almost equal macroblocks using a smart thresholding method. The algorithm is implemented and evaluated as part of an embedded vision system for room surveillance.