Supervised Noise Reduction for Clustering on Automotive 4D Radar

Abstract

In the automotive industry, radar technology is an essential component for object identification due to its low cost and robust accuracy in harsh weather conditions. Clustering, an unsupervised machine learning technique, groups together individual radar responses to detect objects. Because clustering is a significant step in the automotive object identification pipeline, cluster quality and speed are especially critical. To that extent, density-based clustering algorithms have made significant progress due to their ability to operate on data sets with an unknown quantity of clusters. However, many density-based clustering algorithms such as DBSCAN remain unable to deal with inherently noisy radar data. Furthermore, many existing algorithms are not adapted to operate on state-of-the-art 4D radar systems. Thus, we propose a novel pipeline that utilizes supervised machine learning to predict noisy points on 4D radar point clouds by leveraging historical data. We then input noise predictions into two proposed cluster formation approaches, respectively involving dynamic and fixed search radii. Our best performing model performs roughly 153 percent better than the baseline DBSCAN in terms of V-Measure, and our quickest model finishes in 75 percent less time than DBSCAN while performing 130 percent better in terms of V-Measure.