Enhancing Vehicular Navigation in Urban Environments with Radar Point Clouds: A Pre-Filtering and Sensor Fusion Approach

Abstract

Accurate positioning is crucial for the safe operation of autonomous vehicles, particularly in urban environments where Global Navigation Satellite System (GNSS) signals are degraded due to signal blockages, multipath errors, and environmental obstructions. Inertial Navigation Systems (INS) can provide an alternative positioning solution when GNSS is unreliable by integrating measurements from Inertial Measurement Units (IMUs) to determine the vehicle position, velocity, and attitude. However, low-cost and commercial-grade IMUs found in land vehicles suffer from inherent inertial sensor errors, including bias drift and scale factor instability. These errors cause INS to be reliable only over the short term and require external corrections to remain reliable over mid- to long-term GNSS outages. Exteroceptive sensors such as cameras, lidar, and radar can enhance positioning by providing additional environmental references. Among these, automotive radar operating at 77 GHz is advantageous due to its resilience to adverse weather and varying lighting conditions, as well as its ability to provide radar cross-section (RCS) and Doppler velocity measurements. In particular, positioning systems based on registering radar scans to prior maps of the environment have shown promise as an alternative positioning solution. However, radar-based positioning presents challenges, as low-cost automotive radars are susceptible to noise, ghost detections, and a limited number of detections per scan, all factors which can potentially degrade the performance of radar-based positioning algorithms. This study develops radar point cloud filtering techniques designed to remove dynamic objects and noise from radar data, thereby improving the accuracy of radar scan-to-map registration. The point cloud preprocessing techniques developed in this work employ machine learning and classical filtering techniques, such as velocity-based filtering, geometric clustering, and support vector machine (SVM) classification, to enhance the static environment detected by the radar using Doppler, RCS, and positional information. Furthermore, the study evaluates the performance of two state estimation techniques, the Error State Extended Kalman Filter (ES-EKF) and the Unscented Kalman Filter (UKF), across multiple real-world urban driving scenarios, analyzing their robustness and accuracy. Experimental results demonstrate that the proposed filtering approach improves radar-based positioning in urban environments, improving the reliability of autonomous vehicle navigation in Urban and GNSS-denied conditions.