交通视觉中鲁棒目标检测方法研究

With the rapid development of highway and high-speed railway in our country, it becomes more and more important to reduce traffic accidents and ensure transportation safety. Intelligent video surveillance is an important means to provide traffic safety and has a great application prospect. It can effectively detect vehicles, pedestrians, road situation and anomaly in the traffic lines. But one of its key technologies-object detection-has not yet been solved effectively. Object detection still cannot be effectively used in some complex traffic environments like light changes, rain and snow disturbance, camera shake, target blocking and appearance changes. Aiming at the issues of object detection in traffic scene, this project studies: robust detection of moving targets in noise background; robust detection of moving objects in non-stationary background (camera with movement); robust detection of specific object in complex background. Three methods are proposed: moving object detection based on brightness compensation and local kernel histogram; moving object detection based on distributed motion estimation in moving background; specific object detection based on structural description with graph model. On this basis, a robust object detection system platform for complex and changing traffic environment will be designed and implemented to verify our proposed theoretical approaches, promoting the research results to practical applications in our nation's intelligent transportation.

随着我国高速公路与高速铁路的快速建设，对如何减少交通事故，确保运输安全提出了新的挑战。智能视频监控是保障交通安全的重要手段之一，它可有效检测交通线路中的车辆、行人、道路等情况和异常，具有巨大的应用前景，但其中的关键技术目标检测没有得到有效解决，不能有效处理复杂交通环境中的光照变化、雨雪扰动、摄像机摇动、目标遮挡及外观变化等情况下的目标检测。本申请项目针对交通场景目标检测所存在的问题，研究解决干扰背景中运动目标的鲁棒检测、背景运动（摄像机非静止）情况下的运动目标的鲁棒检测、复杂前景中特定目标的鲁棒检测三方面问题，提出了基于亮度补偿与局部核直方图的干扰背景下运动目标检测方法、基于分布式运动估计的运动背景下运动目标检测方法、基于图模型结构化描述的特定目标检测方法，并在此基础上设计实现一个针对复杂多变交通环境的鲁棒目标检测系统平台，验证理论方法，促进研究成果走向实用，应用于我国智能交通。

本项目以交通视觉问题为研究对象,对如何在复杂多变环境下进行鲁棒目标检测展开研究,主要研究了三方面问题: 光照变化等干扰背景中运动目标鲁棒检测问题,背景变化情况下的运动目标鲁棒检测问题,复杂前景中特定目标鲁棒检测问题。通过对此三方面问题研究，发表标注本项目支持的论文60多篇，其中SCI检索期刊论文25篇（包括IEEE Trans论文5篇，另有1篇IEEE Trans论文录用），国际学术会议论文37篇。发明专利方面，共有8项发明专利获授权/新申请。人才培养方面，共有9名博士生毕业取得博士学位，1名博士后出站，20多名硕士生毕业。提出了面向光照的剧烈变化场景中运动目标检测分割的颜色恢复鲁棒前景目标预测算法等解决三方面问题的方法。