Development of an AI-based automatic traffic volume monitoring system using video footage and evaluation of its applicability conditions

Abstract

Traffic volume surveys are fundamental for road planning and maintenance, yet manual observation requires significant labor and cost. This study proposes an AI-based automatic traffic volume observation system using video footage, integrating a camera-shake-robust composite matching score, dual virtual counting lines for directional estimation, and time-series weighted class re-evaluation, built on the YOLOX detector and ByteTrack tracker. A two-phase evaluation was conducted: Phase I (short-duration ablation, 2 min, 18 GT vehicles) reduced count error from 33.3% to 22.2% and eliminated ID switches (3 to 0) with 100% direction accuracy; Phase II (long-duration, 12 hours, approx. 11,000 vehicles) achieved 0.15% error for the upbound direction but 38.0% systematic undercount for the downbound direction due to occlusion. A synthetic-shake simulation further quantified the tolerance limit of the proposed tracker at approximately 30 px (95-percentile). These results demonstrate that camera geometry and inter-vehicle occlusion are the decisive performance factors, and provide quantitative criteria — including a recommended camera-height-to-road-width ratio of 1.0 or greater — for practical deployment decisions.