Leverage Frequency Domain Neural Networks in Parking Lot Occupancy Detection on FPGA Devices

Parking guidance systems have gained significant traction in the development of smart cities. A critical component of these systems is the algorithm that helps drivers locate available parking spaces in desired areas. Traditionally, this task is accomplished using neural network classifiers with camera footage. In this study, we develop a deep learning network model that integrates convolutional computation with Fast Fourier Transform (FFT) techniques to reduce both the number of parameters and the computational load during inference. The objective is to propose an alternative to Convolutional Neural Networks (CNNs) that can minimize edge computation requirements. We implemented a parking application utilizing the proposed neural network on programmable hardware, specifically the Xilinx Zynq UltraScale+ MPSoC, known for its low latency and energy efficiency. Our results indicate that this solution reduces the model parameter size by approximately 32.7% and the inference time by about 8.4% compared to a CNN model with a similar architecture.