Vehicle Counting Robust Against Occlusions Using Newell Model
Yoshitaka Meguro, Yota Yamamoto, Hideki Yaginuma, Yukinobu Taniguchi · IEEE Access · 2026
Accurate traffic-volume measurement is essential for effective urban planning, traffic management, and congestion countermeasures. A cost-efficient way to obtain these measurements is to repurpose existing closed-circuit television (CCTV) cameras in combination with image-based vehicle counting methods. However, conventional methods struggle with accuracy when occlusions are created by other vehicles or road lighting fixtures, leading to detection failures and ID switches. To address this problem, we propose a vehicle counting method robust to occlusions as it utilizes a car-following model. The method predicts the position of occluded vehicles using the car-following model; tracking errors are mitigated even in heavy-occlusion scenarios. We conduct experiments on 5.5 hours of CCTV footage from four roadway locations in Japan. The results demonstrate that the proposed method significantly reduces accuracy degradation even at congested sites with frequent occlusions. Our method achieves a mean absolute percentage error (MAPE) of 3.8% and an identification F1 score (IDF1) of 86.1%. Compared to conventional methods under congested conditions, our method improves MAPE by 31.7 points and IDF1 by 7.0 points, respectively.