Abstract
This paper focuses on the power efficient design on the FPGA SoC for the object detection system based on Binary Convolutional Neural Network (BCNN). Especially, for the small IoT devices, such as an intelligent dash-cam, computer vision system installed on an unmanned aerial vehicle, the power consumption could be a significant factor of the performance and scalability. However, the optimized FPGA design has limitations to reduce the overall power consumption amount. We focus on the design of the FPGA Accelerator as well as the effective design of the peripherals including CPU. In our proposed FPGA-SoC design, it supports not only FPGA but also CPU and the peripheral component can be supported by additional virtual memory system for reducing the processing time. Overall customization including customized BCNN, virtual memories for CPU and FPGA part allows our testbed to achieve low power consumption without speed degradation. Our testbed is based on customized YOLOv2 which consists of applied binary and half precision convolution, and pipeline-based architecture with accelerated hardware design on the target device. The target device used in this paper is the Xilinx ZYNQ-SoC based PYNQ Z-1 board. Our proposed system achieves 15.15 frames per second (FPS) and 1.45 watts of power dissipation. Our result shows that our design technique is effective for real-time object detection and low power system.