Surface Classification with Millimeter-Wave Radar Using Temporal Features and Machine Learning

Classification of surfaces using millimeter-wave radar commonly considers the use of polarization-based methods for road condition monitoring. When a surface consists of larger structures, one is instead often interested in monitoring the surface topography, which is typically not resolvable by the limited radar bandwidth. To alleviate this problem, we here consider several phase coherent radar measurments conducted during the motion of the radar, in order to capture not only the instantaneous depth measurement, but also the depth variation over time. Analysis of scattering from rough surfaces is highly complex and relies on prior knowledge of surface structure. By constructing a set of features based on a number of radar measurements over time, a machine learning classifier is proposed to distinguish grass target surfaces from asphalt, gravel, soil, and tiled surfaces. Six different classifier structures are evauated and presented in the paper. Using estimated autocovariances and average envelope shapes as features, a small, fully connected, neural network classifier is, using a leave-one-out strategy, shown to allow for accurate determination of the surface type. The proposed classifier can be implemented with limited hardware requirements, making it suitable for autonomous devices, such as, e.g., autonomous lawn mowers.