Power Efficient Real-Time Traffic Signal Classification for Autonomous Driving Using FPGAs

The significant advancements in research within the domain of Advanced Driver-Assistance Systems (ADAS) have been significantly propelled by the integration of Artificial Intelligence (AI) and Deep Learning (DL). The detection and classification of traffic signals are pivotal for ensuring driver safety and comfort, prompting the development of numerous detection and classification methodologies in recent times. Convolutional Neural Networks (CNNs) and Field Programmable Gate Arrays (FPGAs) have individually demonstrated notable efficacy in signal processing, image detection, and image classification tasks. However, the amalgamation of CNNs and FPGAs for traffic signal classification offers enhanced flexibility, accuracy, and resource efficiency. This study proposes a software-hardware co-design approach for the development of a resource-efficient real-time traffic signal classification system tailored for FPGAs utilizing the CNN algorithm. The software implementation has been carried out within the TensorFlow framework, while the hardware implementation is executed via the Xilinx Vivado 2023.1 platform and rigorously evaluated using the Xilinx ZCU102 FPGA board. Experimental findings demonstrate that the traffic signal classification accuracy can attain approximately 99.98% at a frame rate of 84139 frames per second (FPS) on the FPGA platform, with a power consumption of 4.4W. The exceptionally low latency and heightened accuracy exhibited by the proposed system render it highly advantageous for ADAS and other safety-critical applications.