FPGAとDeep Learningを用いたADASの構築

抄録

Today, training and execution of deep learning is almost carried out on GPU because of extensive calculation. However, using GPU have problems of power consumption and portability. This paper presents the way to recognize and estimate size of automobile license plate using deep learning on FPGA. Binarized Neural Network to reduce calculation on PYNQ-Z1 board to design of hardware is used for the image processing. As a result, the recognition accuracy of automobile license plate was 92%, and the size estimation accuracy of automobile license plate was 87% within ± 8 pixels. In addition, processing speed of hardware was 1,604 microseconds and it had about five hundred times more than CPU.