高质量的交通数据是分析和解决交通问题、提升交通管理智能化水平的关键。与现有交通数据采集方式相比，无人机具有灵活性强、视野广、成本低等优势。本文的研究目的是：从多种天气和光照条件下采集的无人机航拍视频中提取出高精度的车辆轨迹、速度和航向角等车辆运动数据。研究目标是：针对现有研究存在的问题，开发一种基于无人机航拍视频的车辆运动数据提取框架。本文的研究运用了图像处理、计算机视觉和深度学习等技术，主要研究成果如下：无人机航拍视频采集与预处理方面，本文提出了一种视频稳像算法，采用ORB特征点提取算法和FLANN特征点匹配算法，通过仿射变换将视频帧与第一帧配准，实现了无人机航拍视频稳像。车辆目标检测方面，本文提出了一种基于定向包围框的旋转目标检测算法YOLOv5-OBB。提出了一种与CSL标签相结合的六参数回归方法，可提取车辆的旋转角度和航向信息。引入CLAHE图像增强算法，增强了对多种天气和光照条件下拍摄的无人机航拍图像的车辆检测准确性和鲁棒性。车辆目标跟踪方面，本文提出了一种针对旋转多目标的SORT++跟踪算法。基于旋转目标检测结果和目标的外观特征、运动特征设计了包围框过滤模块、外观特征提取模块和运动特征提取模块，提高了跟踪算法的准确性和鲁棒性。并对卡尔曼滤波算法进行了改进，实现对旋转目标的运动状态进行更新和预测，以获取更准确、更稳定的旋转目标跟踪结果。车辆运动数据提取方面，本文采用了高斯平滑插值算法，首先使用线性插值补齐缺失数据，再使用高斯平滑对提取的车辆运动数据进行平滑降噪，提高了数据精度。此外，本文设计了实地实验，使用车载高精度RTK-GPS传感器采集车辆的差分定位、GPS速度和GPS航向角度数据作为真值，验证了本文设计的车辆运动数据提取框架的准确度和鲁棒性。本文所设计的基于无人机航拍视频的车辆运动数据提取方法研究，具有重要的理论意义，提取的高精度车辆运动数据在交通管理和自动驾驶等方面具有重要的应用价值。
High-quality traffic data is the key to analyzing and solving traffic problems and improving the level of intelligent traffic management. Compared with existing traffic data collection methods, UAVs have significant advantages such as strong flexibility, a wide field of view, and low cost. The purpose of this study is to extract high-precision vehicle trajectories, speeds, yaw angles, and other vehicle motion data from the UAV videos captured under different weather and illumination conditions. And the goal of this work is to propose a framework for extracting vehicle motion data from UAV videos by analyzing the inadequacies of existing research. The research in this work used algorithms for image processing, computer vision, and deep learning, with the main research findings as follows:For UAV video capture and preprocessing, this work proposes a video image stabilization algorithm using the ORB feature point extraction algorithm and the FLANN feature point matching algorithm, which completes UAV video stabilization by registering other video frames with the first frame through affine transformation.For vehicle detection, this work proposes a rotated object detection algorithm using oriented bounding boxes, called YOLOv5-OBB. A six-parameter regression method was proposed and combined with CSL labels to extract vehicle yaw angle and heading information. The added CLAHE image enhancement algorithm further improves the vehicle detection accuracy and robustness of unmanned aerial vehicle images captured under various weather and illumination conditions.For vehicle target tracking, this work proposes the SORT++ tracking algorithm for multiple rotated objects. To increase the accuracy and robustness of object tracking, the bounding box filtering module, appearance feature extraction module, and motion feature extraction module are designed based on the detection results of rotated objects and their appearance and motion features. Besides, the Kalman filtering algorithm is improved to predict and update the motion state of the rotated objects to obtain more accurate and stable tracking results.For vehicle motion data extraction, this work uses the Gaussian smooth interpolation algorithm, which first fills in the missing data by linear interpolation and then smooths and reduces the noise of the extracted vehicle motion data using the Gaussian process regressor to improve the data accuracy. In addition, field experiments were conducted to collect reference data, including differential positioning, GPS speed, and GPS yaw angle data of vehicles from an onboard high-precision RTK-GPS sensor, to evaluate the precision and robustness of the vehicle motion data extraction framework proposed in this work.The research on methods for extracting vehicle motion data from UAV videos designed in this work has important theoretical significance. The extracted high-precision vehicle motion data has important application value in traffic management and autonomous driving.
