智能交通中基于移动视频的目标快速识别方法研究

Targets detection and tracking based on video is very important for safe driving and unmanned vehicles in intelligent transportation systems. This project intends to discuss the methods on targets detection and identification in mobile video. Firstly, how to locate targets under various environmental conditions will be researched. We will analyze the optical flow of different targets in mobile video, and track the feature distribution trajectory in time domain. Secondly, we will research target motion model based on probability. Then we will discuss how to separate and track the targets from background using hidden Markov model. Thirdly, we will research road sign detection method based on image segmentation and shape analysis techniques. Then we will determine the class of the road sign by shape feature, which can reduce the matching computation greatly. And we will match the unknown signs with the known reference road signs stored in the database using support vector machine technology. This project intends to adopt a combination of offline experiments with online experiments. Lastly, the targets detection and identification methods will be transplanted to SBC in order to test the real-time and effectiveness of the system.

基于车载视频的智能交通系统中，对前方车辆、交通标志等目标的实时检测与跟踪对于车辆安全驾驶、无人驾驶等具有重要意义，本课题拟对移动视频中目标的检测与识别方法进行探讨。第一，拟对复杂环境下的目标定位方法进行研究，拟通过分析视频中不同目标运动的光流特征，对在时间域上分布的特征轨迹进行跟踪。第二，拟根据背景运动特征和目标运动特征研究基于视点的运动概率模型，讨论运用隐马尔可夫模型将目标从背景中分离出来的方法，并对其进行概率跟踪。第三，拟研究融合图像聚类分割和形状分析技术的道路交通标志检测方法；并通过分析形状特征确定交通标志的类别，将识别时的匹配范围减少到具有相同形状的参考标志子集，然后拟利用支持向量机技术对待识别标志和参考标志子集内的图像进行相似性度量。本项目拟采用离线实验与在线实验相结合的方式，最终将对移动视频中车辆、背景及交通标志的目标检测、识别方法移植到单板机上，以测试本方法的有效性和实时性。

智能交通中对关键目标的实时检测识别对于车辆安全驾驶，辅助驾驶等具有重要意义，本课题对移动视频中关键目标的检测与识别方法进行了探讨。首先研究了对复杂环境条件下的视频进行目标定位的方法，研究在AdaBoost算法框架下，将Haar-like特征和LBP特征进行特征后融合，并通过对比实验说明了算法在降低错检率、提高检测速度方面的有效性；其次，根据车辆特征提出了基于支持向量机的目标识别方法,通过对四类车型的分类实验证明了算法的鲁棒性；最后提出了一种实用的车道线检测方法，根据交通状况提出了道路拥挤度5级设定，设计系统实现了道路状况的预警、违章车辆自动抓拍等功能。本课题的研究对于提高智能交通系统中关键目标的检测识别速度和准确性有重要意义，并对实际交通管控具有重要应用价值。