Design and Implementation of CNN Traffic Lights Classification Based on FPGA

Abstract

Convolutional Neural Networks (CNN) has acquired remarkable achievements in image classification and target detection. It is of great significance to the field of unmanned driving. As a professional accelerating device that is good at parallel computing, FPGA can fully exploit the parallel processing capacity of CNN. Deploying CNN on Field Programmable Gate Array (FPGA) has a wide range of applications. This paper proposes a tri-classification convolutional neural networks with a simple structure and easy FPGA implementation, which is used to realize traffic lights recognition. Based on the understanding of the convolution process, this design adopts methods such as optimized array structure, multi-level data multiplexing, and positive index classification to improve parallelism in the hardware implementation process, which can take into account speed and performance. Transplant the improved CNN to FPGA through High-level Synthesis (HLS) software. The experimental results present that the accuracy of traffic lights classification can reach about 99%, and the processing speed on the FPGA platform is 22 times higher than that on Inter Core i5-8300H CPU.