A Deep Convolutional Neural Network Based Small Scale Test-bed for Autonomous Car

Abstract

In deep end-to-end learning based autonomous car design, inferencing the signal by trained model is one of the critical issues, particularly, in case of embedded component. Researchers from both academia and industry have been putting their enormous efforts in making this critical autonomous driving more reliable and safer. As research on the real car is costly and poses safety issue, we have developed a small scale, low-cost, deep convolutional neural network powered self-driving car model. Its learning model adopted from NVIDIA's DAVE-2 which is a real autonomous car and Kansas University's small scale DeepPicar. Similar to DAVE-2, its neural architecture uses 5 convolution layer and 3 fully connected layers with 250,000 parameters. We have considered Raspberry Pi 3B+ as the processing platform with Quad-core 1.4 GHz CPU based on A53 architecture which is capable to support CNN learning model.