Exploring Depth Estimation on Intel IPU Platform

Abstract

Numerous applications from autonomous vehicles to surveillance systems can benefit from "seeing in 3D". A crucial element of sight is depth detection since this enables evaluation of position and shape. The depth at which objects are located can be estimated by using two or more cameras and comparing the resulting images. Despite the increasing popularity of these algorithms, available solutions require expensive high-end platforms, which are not suitable for commercial applications. These motivate the search for a cheaper alternative and in this work, we provide a low-cost, low-power embedded solution to determine the depth at high speed. A highly parallel but not inherently complex processor architecture is required for this tasks. VLIW processors are the best choices, and the Intel Image Processing Unit is an excellent option. By adding different functional units to it and adjusting an algorithm to take full advantage of them, a 640x480 image pair with 64 disparities can be processed at 36.75fps, which is an improvement of 23\% compared to the best state-of-the-art image processor.