Impact of sensing modality on surrogate safety metrics: A radar–camera fusion framework for signalized intersection approach monitoring

Abstract

Camera systems are common in surrogate safety analysis at signalized intersections, but their velocity estimates remain underexamined. For approach-lane following, frame-to-frame position differencing introduces errors from bounding-box jitter and perspective distortion that propagate into Time-to-Collision (TTC) and a longitudinal following time gap (FTG). We develop a roadside radar–camera fusion framework combining Doppler measurements with CCTV detections using geometric association and a learned confidence model, evaluated against drone trajectories at two intersections. Unfiltered camera estimation produces 3.26 times the exposure-normalized rate of critical TTC episodes that fusion reports. Temporal filtering reduces noise but compresses leader–follower speed differences, suppressing TTC observations. Fusion reduces velocity error standard deviation from 12.59 to 7.39 mph (41.3%) and raises correlation with drone velocity from $r=0.548$ to $r=0.760$ ($\Delta r=0.212$). Savitzky–Golay and Kalman smoothing close this gap, but they use future observations or require ground-truth tuning. Better velocity correlation does not guarantee better TTC results, because correlation evaluates individual speeds whereas TTC depends on the instantaneous difference between them. Two filtered versions of fusion with nearly identical correlation ($\Delta r=0.0039$) report critical rates 9.74 points apart. The Kalman-smoothed camera, most accurate on velocity, differs from the drone critical TTC rate by 37.72 points. Unfiltered camera estimation flags a critical interaction where the drone does not in 5.60% of pair-frames, against 1.90% for fusion. This pattern replicates directionally at a second site with an oblique camera view. Velocity validation alone cannot establish safety-metric reliability. Radar–camera fusion can provide a more stable basis for monitoring at both sites.