VGC-RIO: A Tightly Integrated Radar-Inertial Odometry with Spatial Weighted Doppler Velocity and Local Geometric Constrained RCS Histograms
Jianguang Xiang, Xiaofeng He, Zizhuo Chen, Lilian Zhang, Xincan Luo, and Jun Mao
AI summary
Problem
Existing 4D radar-inertial odometry methods struggle with sparse, noisy point clouds and uneven spatial distributions, often over-relying on Doppler constraints from dense regions and failing in unstructured or dynamic environments.
Approach
The authors introduce VGC-RIO, a tightly coupled system that fuses IMU pre-integration with spatially adaptive Doppler weighting and a new Local Geometric Constrained (LGC) RCS histogram for robust point-to-point scan matching in a sliding window optimizer.
Key results
- Adaptive spatial weighting prevents Doppler constraints from being dominated by dense point clusters
- Novel LGC histogram enables robust point-to-point registration under noisy and vigorous motion
- Improved localization accuracy and mapping consistency across multiple public and self-constructed datasets
- Enhanced robustness against sensor jitter and dynamic object interference
Why it matters
Delivers a reliable, weather-resilient localization solution for autonomous vehicles and robots operating in GNSS-denied or adverse conditions where traditional sensors fail.
Abstract
Recent advances in 4D radar-inertial odometry have demonstrated promising potential for autonomous localization under adverse conditions. However, effective handling of sparse and noisy radar measurements remains a critical challenge. In this letter, we propose a novel 4D radar-inertial odometry that fuses inertial pre-integration, radar scan matching and radar Doppler velocity in a tight way. Unlike most radar-inertial odometry that fuses the Doppler velocity with equal weights, we integrate each radar point’s Doppler reading with an adaptive method that can adjust the weights according to the non-uniform point distribution. We further design a new point descriptor for point-to-point matching by combining the point cloud’s local geometric and RCS (Radar Cross Section) information in a histogram. Extensive experiments conducted on multiple datasets demonstrate its localization accuracy improvement and adaptability under different environments and motion conditions.