In recent years, a large number of intelligent unmanned aerial vehicles (UAVs) have emerged in people's vision, leading to the development of a new safe and efficient urban air mobility (UAM) system, which is suitable for both manned and cargo scenarios. Driven by the market, this service is expected to become a new trend in the development of civil aviation. The rapid development of deep learning has brought prospects to urban air traffic, but the increasing data volume and model complexity pose challenges in terms of resource consumption and model deployment. A method of applying spiking neural network (SNN) algorithms in UAM is proposed in this paper, utilizing spike-based transmission to significantly reduce energy consumption compared to deep learning algorithms. To address the issue of missing urban air traffic data sets, we build a complex urban air traffic data set comprising 10 categories such as pedestrians, vehicles, buildings, traffic signals, billboards, garbage bins, unmanned aerial vehicles, hot air balloons, cats, and dogs. In the network model, we introduce the biological neurons such as Integrate-and-Fire (IF) and Leaky Integrate-and-Fire (LIF), achieving the highest accuracy of 81.937% and 82.772%, respectively. Based on the LIF neuron, we propose the Self-Learning Leaky Integrate-and-Fire (SLLIF) neuron, which autonomously learns stimulus-input ratio relationships to better align with the brain's automatic optimization mechanism, achieving a recognition accuracy of 85.484%. Furthermore, we re-evaluate the selection of the hyper parameter time steps and choose suitable values for different neural network models to reduce resource consumption while maximizing accuracy.
|