Enhancing Vulnerable Road User Classification Through Micro-Doppler and Deep Learning: The Impact of Time Window

Abstract

Recent developments in driving technology have led to the creation of advanced driver assistance systems and progress towards fully autonomous vehicles. Cars equipped with radar technology can simultaneously detect multiple vulnerable road users, assessing their distance, speed, and approach angle. For autonomous vehicles to be deemed safe for public roads, they must effectively identify and classify these users. This study employs time–frequency analysis and deep learning techniques to classify spectrograms derived from targets. The training and testing datasets were generated using frequency-modulated continuous-wave (FMCW) radar signals operating at 77 GHz. A five-layer convolutional neural network (CNN) was trained for this purpose. We investigated how different time window types and durations affect the Short-Time Fourier Transform calculation and the CNN classification accuracy for each scenario. As the length of the time window increases, frequency resolution improves, enabling better differentiation between closely spaced frequencies and enhancing classification accuracy. However, increased time window lengths lead to decreased time resolution, causing accuracy to plateau at 800; beyond this point, accuracy declines. We achieved an accuracy rate of 88.95% in classifying seven data classes, with improvements in specific classes compared to prior studies. The findings suggest that micro-Doppler-based convolutional neural networks can effectively classify vulnerable road users, contributing to collision avoidance efforts.