Driving Risk Recognition from Dashcam Video via License-Plate Bounding-Box Scale Variation

Abstract

Dashcams are widely deployed, yet fleet management and insurance risk control still depend on manual frame-by-frame replay for batch retrospective review, while retrofitted recorders offer neither calibration conditions nor real-time links. This paper proposes a calibration-free method that recognizes driving risk from recorded dashcam video through license-plate bounding-box scale variation. Frames are sampled at 5 fps and processed by HyperLPR3 detection and recognition with a confidence threshold of 0.65. Taking the statutory GA 36 small-vehicle plate width of 440 mm as a prior, distance is estimated as D=f×0.44/w, and the scale change rate drives four types of risk rules gated by phase correlation-based stationary suppression. Per-character confidence-weighted voting, three-tier confidence grading, and a manual review loop secure plate identities, and cross-video same-vehicle merging operates within a −30 s~+150 s temporal window. On 35 midday real-world videos totaling 2098.2 s, the pipeline screened out 41 high-risk events (21 lateral cut-in, 18 rapid approach, 1 too-close following, 1 adjacent-vehicle proximity) with a median minimum distance of 5.21 m; 165 trajectories were merged into 118 vehicles, and 38 of 86 medium- and low-confidence plates were corrected in review. The results suggest that the pipeline supports batch, calibration-free, relative-tier screening. The detected events are heuristic screening warnings, provided solely as a reference for post-event retrospective review, and do not constitute a basis for accident liability determination.