Recently, ego-motion estimation using auxiliary sensors has been actively studied in autonomous vehicles, mobile robots, and integrated navigation systems to provide reliable motion states in environments where Global Navigation Satellite System (GN...
Recently, ego-motion estimation using auxiliary sensors has been actively studied in autonomous vehicles, mobile robots, and integrated navigation systems to provide reliable motion states in environments where Global Navigation Satellite System (GNSS) signals are degraded or unavailable. Among these motion states, ego-velocity is directly used in state estimation, dead reckoning, path planning, and control. Biases or temporal discontinuities in the estimated velocity can accumulate as position errors through time integration and degrade navigation stability.
Radar has attracted attention as a navigation sensor because it is robust to adverse environmental conditions and directly provides Doppler velocity measurements. However, radar-based ego-velocity estimation in urban environments remains challenging due to dynamic objects, multipath reflections, ghost targets, clutter, and sparse radar point clouds. Conventional radar-based methods generally estimate ego-velocity after selecting static points from radar observations. When static point selection becomes unstable, the estimated velocity can be biased, divergent, or temporally discontinuous.
To address these limitations, this thesis proposes a robust ego-velocity estimation method based on point-wise reliability weighting and radar-camera fusion. The radar-based proposed method assigns continuous reliability weights to individual radar points using radar cross-section information and velocity-space clustering consistency. These weights are reflected in a Huber loss-based Iteratively Reweighted Least Squares (IRLS) optimization framework to reduce the influence of dynamic objects and outlying observations without relying on hard static-dynamic point selection.
The proposed method is further extended by incorporating camera image information. Radar points are projected onto the camera image plane, and object detection results with projection uncertainty are used to evaluate the spatial association between radar points and detected objects. An object-class-based prior reliability is then combined with the physical consistency of Doppler residuals to compute camera-based point-wise reliability weights. The robust velocity measurement generated at each frame is finally combined with a Kalman filter-based state prediction and measurement update structure to improve temporal continuity.
The proposed method was evaluated on the View of Delft and MSC-RAD4R datasets and compared with RANSAC+LSQ, EFEAR-4D, REVE, and RAVE. On the View of Delft dataset, the radar-based proposed method achieved a Speed RMSE of 0.1236 m/s and an Absolute Trajectory Error (ATE) of 0.511 m. The radar-camera fusion method further improved the result to a Speed RMSE of 0.1234 m/s and an ATE of 0.474 m. On the MSC-RAD4R dataset, the radar-based proposed method achieved a Speed RMSE of 0.4922 m/s and an ATE of 8.163 m, while the radar-camera fusion method reduced them to 0.2413 m/s and 0.885 m, respectively.
These results show that the proposed method improves both single-frame ego-velocity estimation and dead reckoning stability. In particular, the radar-camera fusion structure effectively reduces performance degradation in dynamic urban scenes where radar-only information is insufficient. Therefore, this thesis demonstrates that radar Doppler-based ego-velocity estimation combined with point-wise reliability weighting, robust optimization, and camera-based semantic information can serve as a reliable navigation component in GNSS-degraded environments.