The vision-based traffic flow parameter estimation is a challenging problem especially for dense traffic scenes, due to the difficulties of occlusion, small-size and dense traffic etc. Yet, previous methods mainly use detection and tracking methods to do vehicle counting in non-dense traffic scenes and few of them further estimate traffic flow parameters in dense traffic scenes. A framework is proposed to count vehicles and estimate traffic flow parameters in dense traffic scenes. First, a pyramid-YOLO network is proposed for detecting vehicles in dense scenes, which can effectively detect small-size and occluded vehicles. Second, the authors design a line of interest counting method based on restricted multi-tracking, which counts vehicles crossing a counting line at a certain time duration. The proposed tracking method tracks short-term vehicle trajectories near the counting line and analyses the trajectories, thus improving tracking and counting accuracy. Third, based on the detection and counting results, an estimation model is proposed to estimate traffic flow parameters of volume, speed and density. The evaluation experiments on the databases with dense traffic scenes show that the proposed framework can efficiently count vehicles and estimate traffic flow parameters with high accuracy and outperforms the representative estimation methods in comparison.

中文翻译：

---

密集交通场景下的车辆计数和交通流参数估计

基于视觉的交通流参数估计是一个具有挑战性的问题，特别是对于拥挤的交通场景，由于遮挡，小尺寸和拥挤的交通等困难。密集交通场景，其中很少有人进一步估计密集交通场景中的交通流参数。提出了一种在拥挤的交通场景中对车辆进行计数和估计交通流参数的框架。首先，提出了一种金字塔YOLO网络来检测密集场景中的车辆，该网络可以有效地检测小型和被遮挡的车辆。其次，作者设计了一种基于受限多重跟踪的兴趣线计数方法，该方法可对在特定时间段内穿过计数线的车辆进行计数。所提出的跟踪方法跟踪靠近计数线的短期车辆轨迹并分析该轨迹，从而提高了跟踪和计数精度。第三，基于检测和计数结果，提出了一种估计模型，以估计交通流量的数量，速度和密度。在交通场景密集的数据库上进行的评估实验表明，所提出的框架能够高效地对车辆进行计数，并能以较高的准确度估算交通流参数，并且优于具有代表性的估算方法。