Radar–Video Fusion for Tiny Rockfall Detection and Tracking in Complex Field Environments

Small rockfalls are difficult to localize in wide-view slope video because their image support is weak and intermittent during rapid motion, overlap, collision, and occlusion. This study presents a fixed-site radar–video fusion workflow organized into sensing, association, and fusion layers. The sensing layer aligns radar tracks sampled at 1 hertz (Hz) with video at 30 frames per second (fps) and projects each radar track as an uncertainty-guided image search region. The association layer forms compact motion tracklets, filters candidate pairs by spatial, range, azimuth, and motion-memory consistency, and performs gated one-to-one global assignment. The fusion layer maintains a sensor-independent fusion identifier (Fusion ID) through confirmation, bounded coasting, and reacquisition, while retaining the broad radar rectangle as an intermediate diagnostic output. Evaluation used five field sequences totaling 219 s, with 22 physical trajectories and 114 radar-aligned target instances. At the primary 20-pixel criterion, precision, recall, F1 score, and identity F1 score (IDF1) were 0.809, 0.667, 0.731, and 0.702, respectively; the 76 accepted matches had a mean localization error of 3.67 pixels. These results characterize the complete fusion interface and output chain at the recorded site.