A Performance Per Power Efficient Object Detector on an FPGA for Robot Operating System (ROS)

摘要

In recent years, interest in service robots that human support in living spaces such as homes and hospitals is increasing. Service robots demands for an intelligent processing of images and sounds for smooth communication between a robot and a human working. These processes are computationally intensive, and real-time processing is often difficult in software processing by the CPU. Since it is necessary to integrate various hardware and software into the robot to operate it, to make development more efficient, utilization of open source construction of a rapid prototyping environment is indispensable. In the paper, to accelerate the intelligent processing, we use an FPGA as a hardware accelerator on a robot. Also, in order to improve development efficiency, we use a robot operating system (ROS). We implemented YOLOv2 (You Only Look Once version 2) which is one type of object recognition. Compared with a CPU and a GPU, an FPGA based accelerator was superior in power performance efficiency.