项目名称: 复杂环境下交通视频分析的若干关键技术研究
项目编号: No.61304200
项目类型: 青年科学基金项目
立项/批准年度: 2014
项目学科: 自动化技术、计算机技术
项目作者: 王坤峰
作者单位: 中国科学院自动化研究所
项目金额: 23万元
中文摘要: 智能视频分析是从海量视频中自动获取有价值信息的重要技术。由于交通环境具有极高的复杂性,表现在昼夜更替、光照变化、不良天气、目标遮挡等方面,许多已有的视频分析方法在应用时缺乏鲁棒性。本项目将研究复杂环境下交通视频分析的若干关键技术:首先从环境自适应的角度改进虚拟线圈车辆检测技术,通过提取多种强表征能力的图像特征,并利用半监督学习机制在线优化模式分类器,提高该技术对复杂环境的自适应能力;然后利用动态条件随机场模型研究区域层面的目标跟踪,用单个概率图模型联合建模前景标记和目标标记任务,融合利用图像多特征和上下文信息,实现联合的前景分割和目标跟踪;最后结合目标的尺寸和颜色特征,构建级联分类器--包括尺寸特征的判别函数和颜色特征的k近邻分类器,并将半监督学习机制融入k近邻分类器,将提供一种有效的出租车识别方法。本项目通过研究上述基础性关键技术,有利于提高交通视频分析技术在复杂环境下的精度和鲁棒性。
中文关键词: 交通视频分析;复杂环境;车辆检测;目标跟踪;出租车识别
英文摘要: Intelligent video analysis is an important technology used for automatically acquiring valuable information from massive videos. Since the traffic environment has a high level of complexity, including the alternation of day and night, illumination changes, adverse weather, and object occlusion, many existing video analysis methods have a lack of robustness when applied. In this project, we will study several key techniques of traffic video analysis in complex environments. First, we improve the virtual-loop vehicle detection technique from the point of environment self-adaptation. In order to enhance the adaptability of this technique when confronted with complex environments, we extract multiple image features with strong descriptive abilities and use the semi-supervised learning mechanism for online optimization of the pattern classifier. Then, we study the region-level object tracking technique by using dynamic conditional random fields. We jointly model the foreground labeling task and the object labeling task with a single probabilistic graphical model, and fuse multiple image features and context information to achieve joint foreground segmentation and object tracking. Finally, we combine the size and color characteristics of objects to design an effective approach to taxi recognition. We build a cascade c
英文关键词: Traffic video analysis;Complex environments;Vehicle detection;Object tracking;Taxi recognition