Real-Time Crosswalk Segmentation Using FPGA Based Spiking Neural Networks: A Comparison of Integrate-and-Fire and Izhikevich Hardware Neural Models

This paper presents an FPGA-based framework for real-time crosswalk segmentation, employing Spiking Neural Networks (SNNs) and comparing the Integrate-and-Fire (IF) and Izhikevich neuron models. Crosswalk segmentation is crucial in autonomous driving and smart city applications for ensuring pedestrian safety and optimizing traffic flow. While traditional Convolutional Neural Networks (CNNs) deliver high accuracy, they are often unsuitable for resource-constrained real-time systems due to high power demands. In contrast, SNNs on FPGA platforms offer energy-efficient alternatives, suitable for embedded systems. Our findings reveal that the Izhikevich neuron model achieves a segmentation accuracy of 74.6% with a power consumption of 0.726 W, whereas the IF model reaches 62.86% accuracy at a lower power consumption of 0.287 W. This analysis highlights a trade-off between segmentation accuracy and energy efficiency, which may guide the selection of neuron models based on specific application needs in autonomous systems. Future research will explore further optimizations to balance these factors.