Abstract
Traditional 3D radar can only detect the planar characteristic information of a target. Thus, it cannot describe its spatial geometric characteristics, which is critical for accurate vehicle classification. To overcome these limitations, this paper investigates elevation features using 4D millimeter-wave radar data and presents a maximum likelihood estimation (MLE)-based vehicle classification method. The elevation data collected by 4D radar from a real road scenario are applied for further analysis. By establishing radar coordinate systems and geodetic coordinate systems, the distribution feature of vehicles' elevation is analyzed by spatial geometric transformation referring to the radar installation parameters, and a Gaussian-based probability distribution model is subsequently proposed. Further, the data-driven parameter optimization on likelihood probabilities of different vehicle samples is performed using a large-scale elevation dataset, and an MLE-based vehicle classification method is presented for identifying small and large vehicles. The experimental results show that there are significant differences in elevation distribution from two different vehicle types, where large vehicles exhibit a wider range of left-skewed distribution in different cross-sections, while small vehicles are more concentrated with a right-skewed distribution. The Gaussian-based MLE method achieves an accuracy of 92%, precision of 87% and recall of 98%, demonstrating excellent performance for traffic monitoring and related applications.