Developing a custom DSP for vision based human computer interaction applications

Abstract

As the computing power of modern devices become greater, computer vision is increasingly adopted as the means of human-computer interaction. The industry is struggling hard to bring computer vision into the mobile domain, but there are some difficulties mainly due to the computationally expansive nature of vision applications. On the other hand, there are various kinds of vision applications, which means using fixed hardware block will meet the performance criteria, but it will be possible to run only one application. To meet our needs, we can use another programmable component that is specialized for parallel processing and has domain specific instructions. In this paper, we introduce our work on the development of a low-power VLIW processor with vision-application specific parallel computation logic and a set of application-specific instructions, to improve the processor's performance while reducing the power consumption. Our VLIW processor accelerates by adding special instructions on the application's specific part in which processes heavy computation. We introduce our work on developing a custom DSP for vision system that includes designing the processor with Synopsys Processor Designer tool, and porting LLVM compiler. As the target, we ported a hand recognition application to our system, which we have constructed on Xilinx Zynq evaluation board. Even though our DSP was run at much lower frequency, it could run compute intensive part of the application four times faster than ARM processor, greatly speeding up the whole application.