WUT-NGSIM: A High-Precision and Trustworthy Vehicle Trajectory Dataset

The extraction and construction of high-precision and long-distance vehicle trajectory data and microscopic traffic flow characteristics are critical for traffic safety studies. Current research typically relies on a limited number of datasets which suffer from vehicle detection inaccuracy and limitation of the coverage area. Therefore, we establish a high-precision and long-distance vehicle trajectory dataset of urban scenarios, which is also named as WUT-NGSIM. Primary features of the established dataset: (1) The trajectory data are extracted based on a trajectory extraction framework that contains the video stabilization, vehicle detection and tracking,long-distance trajectory construction, trajectory repair and smoothing, motion feature extraction. (2) The broken trajectory data caused by occlusion are connected through fusing vehicle tracking results at different frame rates based on the Kalman filter and Hungarian Algorithm. (2) Long-distance trajectory data from multi-videos are constructed based on video stitching, video fusion and trajectory fusion. The trajectory precision can be guaranteed by reason of the use of deep learning model and the manual error correction. Finally, the vehicle detection accuracy can reach 100% and the length of the trajectory can reach to 620 meters. This trajectory dataset can provide the high-precision and long-distance vehicle trajectory data for those data-driven studies in transportation.

高精度与长距离车辆轨迹数据及其微观交通流特征之提取与构建，对于交通安全研究至关重要。现有研究往往依赖于数量有限的、存在车辆检测不准确及覆盖区域限制的数据集。因此，本研究构建了一个高精度与长距离的车辆轨迹数据集，该数据集亦命名为WUT-NGSIM。该数据集的主要特点如下：（1）轨迹数据基于包含视频稳定化、车辆检测与跟踪、长距离轨迹构建、轨迹修复与平滑、运动特征提取的轨迹提取框架进行提取。（2）通过融合不同帧率下的车辆跟踪结果，并基于卡尔曼滤波和匈牙利算法，连接由遮挡引起的断续轨迹数据。（3）基于视频拼接、视频融合及轨迹融合，构建多视频中的长距离轨迹数据。得益于深度学习模型的使用及人工误差校正，轨迹精度得以保证。最终，车辆检测的准确率可达100%，轨迹长度可达620米。本轨迹数据集可为交通运输领域的数据驱动研究提供高精度与长距离的车辆轨迹数据。

搜集汇总

背景与挑战

背景概述

WUT-NGSIM是一个高精度、长距离的车辆轨迹数据集，专为交通安全性研究设计。它通过深度学习模型和人工纠错确保轨迹精度，车辆检测准确率达到100%，轨迹长度可达620米，解决了现有数据集中车辆检测不准确和覆盖范围有限的问题。