Feature-Level Sensor Fusion Algorithm Using Vision-Based Regions of Interest and Radar Point Clouds

This paper proposes a radar–vision-based feature-level sensor fusion algorithm that operates robustly under varying illumination conditions. In prior work using decision-level sensor fusion, the object state (position and velocity) was estimated separately by the radar and vision sensors, and these estimates were then fused. In contrast, in feature-level sensor fusion, the three-dimensional point cloud from the radar sensor and the region of interest from the vision sensor are first aligned to identify objects, after which a Kalman filter is applied to estimate the object state (position and velocity). Experiments were conducted under vertical and inclined conditions, as well as in bright and dark environments. Compared with the decision-level fusion algorithm, the proposed feature-level fusion algorithm demonstrated improved position estimation accuracy, slightly reduced velocity estimation performance, and approximately three times faster execution. These results validate that leveraging the complementary characteristics of sensors enables reliable object tracking even in challenging environments.