自适应学习的多摄像机目标跟踪

This project orients at security applications in wide-field intelligent video surveillance. We propose an adaptive learning method based multi-camera object tracking theory and application system, which does not need offline classifier learning. Most existing multi-camera object tracking algorithms are based on color features that are illumination sensitive, and methods based on classifier learning requires sample labeling and classifier learning in multi cameras, which is hard to generalize to other multi-camera environments. In this project, we address multi-camera object tracking in disjoint views. We propose illumination invarient local texture feature for object matching, propose an online Gentle AdaBoost algorithm for online feature selection and classifier learning, propose several effective features and its adaptive learning method for scene constraint. We also propose an online object frame sequence selection method and sequence based object matching and association method for robust multi-camera object tracking. Based on the research of this project, we aim to solve three key problems in multi-camera object tracking: the problem of effective feature description, large-scale data labeling, and model generalizaiton and reusability. Therefore, research in this project will advance both research and applications in multi-camera object tracking.

本项目面向安防应用需求，立足于广视域智能视频监控，提出一套自适应学习、无需人工组织训练的多摄像机目标跟踪理论和应用系统。已有的多摄像机目标跟踪算法大都依赖于受光照变化影响较大的颜色特征，并且基于模式分类的方法需要针对多摄像机网络事先组织样本标定和分类器学习，缺乏大规模应用的普适性。本项目针对无重叠视域的多摄像机目标跟踪，提出光照不变的局部纹理特征用于目标匹配，提出在线Gentle AdaBoost算法用于在线特征选择和分类器学习，提出多种有效的环境辅助约束特征及其自适应学习方法用于学习多摄像机场景约束，并提出动态目标序列选择方法和基于序列匹配的多目标关联方法实现可靠的多摄像机目标跟踪。本项目的研究将解决多摄像机目标跟踪中有效特征描述的困难、大量数据标定的困难、以及模型适用性和重用性的困难三大关键问题，从而为多摄像机目标跟踪的实际应用打下基础，推动学术研究和智能安防应用的发展。

本项目面向公共安全应用需求，立足于广视域智能视频监控和目标分析，提出一套多摄像机环境下的鲁棒特征描述方法和基于度量学习的跨摄像机目标再辨识理论和应用系统。首先，我们根据多摄像机环境下目标表象变化的特点，提出几种鲁棒的特征描述方法，包括局部极大出现率特征，显著颜色名特征，Gabor三元模式特征，和基于深度学习的CNN特征。其次，我们提出了跨视域二次判别分析和加权半正定约束的度量学习方法用于跨摄像机目标再辨识。针对多摄像机环境下监控视频目标分析的实际需求，我们还提出了基于多标签CNN的监控目标属性分析方法，多摄像机目标轨迹挖掘方法，和多摄像机联合视频浓缩方法。本项目执行期间还采集了一个多摄像机行人数据库，并作了目标分割和标定，制定了算法评估协议并进行了对比评估。该数据库现已开放获取，以推动该领域的研究。总体而言，本项目的研究积累了多摄像机环境下监控视频目标分析的研究和应用经验，形成了原创算法和海量视频目标检索应用系统，为学术研究和公安视频侦查技术应用均起到有意义的借鉴。