Robust 4D millimeter-wave radar localization in adverse environments: a case study of smoke conditions

Abstract Accurate localization in adverse environments represents one of the key technical challenges for autonomous driving systems. Traditional LiDAR and vision-based localization methods suffer from severe performance degradation or even complete failure in atmospheric scattering conditions such as smoke and dense fog due to optical limitations. To address this challenge, this paper proposes a localization method based on Doppler velocity constraints from 4D millimeter-wave radar. The method fully exploits the all-weather penetration capability of millimeter-wave radar and the rich kinematic information from Doppler measurements. A two-stage architecture is constructed, combining Doppler-enhanced front-end processing with back-end factor graph optimization. Specifically, by establishing a Doppler consistency model to identify and suppress dynamic points that violate rigid body motion assumptions, and designing specialized Doppler-kinematic factors integrated with IMU constraints within a factor graph framework, the method achieves robust global trajectory optimization. In the front-end processing, an adaptive dynamic point suppression algorithm and an improved iterative closest point (ICP) algorithm with joint Doppler-signal to noise ratio weighting are designed, significantly improving the accuracy and robustness of point cloud registration. In the back-end optimization, the method integrates front-end odometry constraints, Doppler velocity constraints, and IMU kinematic constraints, achieving globally consistent trajectory estimation. Experimental results demonstrate that on two smoke datasets, the proposed method achieves mean absolute trajectory errors in the 2D plane of 1.289 m and 1.606 m, respectively. This represents improvements of 92.9% and 91.4% over the 4D Radar ICP baseline method, 42.8% and 60.2% over LiDAR-based localization, and 98.8% and 97.1% over vision-based localization. Vertical elevation errors of 0.408 m and 0.456 m are also achieved. This research provides crucial technical support for safe navigation of autonomous vehicles under adverse weather conditions and advances the development of all-weather autonomous driving technology.