Top

Road Work-Zone Capacity Prediction Under UAV Traffic Video and Detector Data Using an Improved SAM2-TinyViT Model

Authors

Damian Przybylski
80-233, Poland

Files

Abstract

Road work-zone capacity prediction is required for lane-closure planning, queue warning, temporary speed-control review and construction-zone management, but the problem is difficult because capacity is jointly affected by work-zone geometry, lane occupation, heavy-vehicle ratio, merging behaviour, weather visibility, detector reliability and UAV viewing conditions. This paper proposes an improved SAM2-TinyViT model for road work-zone capacity prediction using UAV traffic video and detector data. UAV frames are segmented into open-lane masks, closed-lane masks, drivable regions, queue areas and construction-activity zones by a SAM2-based visual branch. Detector sequences provide flow, speed, occupancy, headway, density and discharge-response descriptors, while source reliability flags indicate viewpoint shift, occlusion, detector dropout and timestamp uncertainty. TinyViT converts mask-guided regions into lightweight visual tokens, and a fusion module integrates spatial bottleneck evidence with detector-based temporal response. The model outputs short-term work-zone capacity, congestion-state class, bottleneck evidence, confidence and recommended management review. Experiments use a simulated multi-work-zone dataset covering lane closure, shoulder occupation, moving maintenance, temporary speed control, heavy-vehicle mixing, queue spillback, detector dropout and unseen-site testing. The framework is designed as traffic-manager decision support for queue warning, speed-control review, lane-control review and manual inspection, rather than a direct automatic control command.

Details