Automatic Driving of End-to-end Convolutional Neural Network Based on MobileNet-V2 Migration Learning

Abstract

Convolutional neural network is gradually mature, followed by the arrival of 5G era, autonomous driving will become a development hotspot. MobileNet is a Convolutional Neural Network used depthwise separable convolutions to decrease parameters so that the devices with limited resources can use it to complete image recognition. In this paper, we use MobileNet-V2 migration learning improvement to simulate automatic driving steering on embedded devices. In this experiment, in our data set, we compared Nvidia end-to-end automated driving network with our migration learning neural network based on MobileNet-V2 end-to-end convolution. The improved MobileNet-V2 network can works on raspberries pi only has CPU faster and real-time prediction to keep in the lane line, ensure the model to reduce the number of parameter at the same time, the identification error decreases.