FPGA-Based DNN Implementation for the Autonomous Car System

Lane detection is one of the most fundamental problems in the wide field of autonomous vehicles. This topic has promoted lots of projects in the academic environment. This chapter describes the development of an autonomous-driving car using a field-programmable gate arrays (FPGA) with convolutional neural network (CNN) hardware. First, we proposed a compact-size CNN model that has successfully tested on FPGA with the TuSimple dataset, the QuantLaneNet. The QuantLaneNet is a low-power design that can detect lanes with high-speed processing. Second, we prepared a new dataset on our university campus and achieved an accuracy of 92.57% with an average 624 frames per second (FPS) (FPGA only). Third, we write a program, the steering angle algorithm, to work with the extracted data from the model and control the vehicle. Since QuantLaneNet only uses the peripheral component interconnect express (PCI-E) protocol for transmission, which causes incompatibility between devices, we alternate it with Ethernet. Using an example design provided by Xilinx, we implemented a design connecting FPGA and personal computer (PC) via Ethernet and transmitting images back and forth. The steering angle program is tested on video using FPGA and PC, with an average FPS of 595.