Multiple-output network for simultaneous target classification and moving direction estimation in automotive radar systems

Abstract

To ensure the stable operation of autonomous driving, a rich understanding of road conditions is indispensable. Therefore, in this paper, we propose a method that uses a single deep learning (DL) network structure to classify commonly encountered on-road objects, such as vehicles, cyclists, and pedestrians, and simultaneously estimating their moving direction. First, we obtain information about the target, such as range, velocity, azimuth angle, and elevation angle, using a four-dimensional imaging radar. Next, we convert the detection results into point cloud data and represent them within a three-dimensional spatial coordinate system. Then, the point cloud data is projected onto the XY-plane to classify the target's class and estimate the moving direction of the target. In the XY-plane, we apply a preprocessing step using density-based spatial clustering to remove noise from the detection results, cluster the targets, and convert this processed data into image data. Subsequently, we conduct training on a multiple-output DL network designed to simultaneously perform object classification and predict the moving direction using the image data. Finally, the performance evaluation of the proposed method results in an object classification accuracy of 96.10%, and the root mean square error for estimated moving directions is 5.54 degrees, 3.89 degrees, and 15.35 degrees for vehicles, cyclists, and pedestrians, respectively, with a runtime of 0.1 s, demonstrating its efficiency.