Artificial Intelligence Design on Embedded Board with Edge Computing for Vehicle Applications

Abstract

This article proposes advanced driver assistance system (ADAS) from neural network by YOLO v3-tiny on vehicle platform of NXP S32V234 with edge computing to detect pedestrians and knights. The implemented embedded board has limitation to perform a lot of convolution. As proposed design need to reduce the amount of operation, the considered problem of reduced precision at the same time. The proposed architecture uses method of Squeeze Net and quantization to reduce the amount of operation about 46% and the precision has only slightly reduced. The proposed methods of image to column (Im2col) and memory efficient convolution (MEC) rearranges continuous matrix space to access. The proposed hardware of APEX uses to accelerate operations can reduce execution time and increase detection speed by ten multiples compared with YOLO v3-tiny architecture.