交通场景下基于视频的智能监控分析关键技术研究

With the development of society and economy, the range of traffic monitoring is expanding. As a result, traditional manual monitoring approaches no longer meet the needs of the current traffic monitoring systems. Integrating video analysis methods, intelligent traffic surveillance technology employs computers to complete the various monitoring tasks and is the inevitable trend of the development of traffic monitoring. In this project, we focus on roadside camera calibration problem and 3D-2D vehicle matching, which are two key problems in intelligent traffic surveillance. Considering that in the existing roadside camera calibration methods the number of the parameters to be estimated is relatively small or the used calibration condition is too strict to satisfy, we will present two solutions: the calibration method based on vanishing point and vanishing line and the calibration method based on vehicle information. In the first solution, we will present a detector of the minimum calibration condition which consists of two vanishing points and a vanishing line, and then apply iterative reweighted least squares to the minimum-calibration-condition-based calibration method to improve its practicality and accuracy. In the second solution, vehicle model is introduced as a new calibration pattern and the camera calibration problem is converted into a vehicle matching problem. Therefore, the vehicle-information-based calibration method is suitable for a variety of traffic scenarios. Considering that the existing vehicle matching methods often poorly perform when the clutter or occlusion exists, we will propose a new 3D-2D vehicle matching method that takes into account both local and global characteristics of vehicle matching. This project is to provide a solid theoretical foundation and an effective technical support for intelligent traffic monitoring.

随着交通监控范围的扩大，传统的人工监控方法已经不能满足当前交通监控的需求。智能交通监控利用计算机完成各种监控任务，是交通监控发展的必然趋势。本课题重点研究智能交通监控的两个关键问题：道路相机标定和车辆3D-2D匹配。针对现有的道路相机标定方法存在所能估计的参数个数少或标定条件苛刻的问题，拟研究两种解决方案：基于消失点和消失线的标定方法和基于车辆信息的标定方法。第一种方案中引入水平消失线到标定条件中，与两个消失点组成最小标定条件，并将迭代重加权策略应用到基于最小标定条件的标定方法中，以提高标定方法的实用性和准确性；第二种方案中引入车辆作为标定物，将道路相机标定问题转化为车辆匹配问题，适用于各种交通场景。针对现有的车辆3D-2D匹配方法易受遮挡、干扰影响的问题，拟研究一种融合局部特性和全局特性的车辆3D-2D匹配方法。本课题的开展将为智能交通监控的研究提供坚实的理论基础和有效的技术支撑。

随着交通监控范围不断扩大，传统人工监控方法的局限性日益明显，智能交通监控已经成为交通监控发展的必然趋势。本项目重点研究了智能交通监控中两个关键问题：道路相机标定和车辆3D-2D匹配。对于道路相机标定问题，提出了两种标定方法，一种是基于两个消失点的动态标定方法，一种是基于车辆3D模型的标定方法。基于两个消失点的动态标定方法利用了车道方向消失点和垂直方向消失点的多个观测值来动态地更新相机参数，通过最大后验概率估计将相机标定问题转化成最小二乘优化问题。该标定方法的特点是放宽了对标定条件的要求，适用于更多交通场景，而且对消失点噪声更加鲁棒，标定结果更加准确。基于车辆3D模型的标定方法首次引入车辆3D模型作为标定物，借鉴传统基于标定物的标定方法，利用车辆3D-2D匹配实现道路相机的标定，将道路相机标定问题转换成为车辆3D-2D匹配问题。该标定方法的特点是通用性强，不依赖于交通场景的几何结构信息，只依赖于车辆信息。对于车辆3D-2D匹配问题，提出了一种基于局部特性和全局特性融合的车辆匹配方法，该匹配方法充分利用了像素梯度信息和车辆剪影信息，它首先改进了基于像素梯度信息的局部匹配，然后将局部匹配和基于车辆剪影信息的全局匹配有机地结合起来。该车辆匹配方法的特点是，和现有的车辆匹配方法相比，不仅能够更加准确地反映车辆匹配程度，而且对遮挡和干扰更加鲁棒。此外，我们将该车辆匹配方法应用于车辆3D定位系统，提出了一种3D姿态估计器，用于估计车辆在路面上的空间位置和方向。上述的研究工作将为智能交通监控的研究提供坚实的理论基础和有力的技术支撑。