项目名称: 弱标注下基于主动学习的检测器适应问题研究
项目编号: No.61202234
项目类型: 青年科学基金项目
立项/批准年度: 2013
项目学科: 计算机科学学科
项目作者: 庞俊彪
作者单位: 北京工业大学
项目金额: 24万元
中文摘要: 在智能视觉监控中,检测器适应技术是保证其在新环境下性能的有效手段。本项目面向智能视觉监控的实际需求,以避免重新训练检测器和减少新环境下样本的人工标注为目标,利用主动学习,开展弱标注下检测器适应问题的研究,包括:多个不可靠信息源下,数据关联的原则与技术;多信息源下的主动标注技术;弱标注下的迁移学习技术等问题。针对视觉监控数据的特点,我们提出了基于多实例学习的领域适应算法、融合协同训练的主动标注算法、基于图模型的多个不可靠信息源下的分类算法。通过本研究的实施,将对智能视觉监控技术起到推动作用,为视频监控服务提供核心算法与技术。
中文关键词: 智能视觉监控;检测器适应;弱标注;特征表示;数据关联
英文摘要: In smart visual surveillance, detector adaptation is an efficient way to guarantee its generalization ability across new environments. Considering the practical applications of smart visual surveillance, this proposal adopts the idea of active learning to research detector adaptation with weak annotated data, towards avoiding re-training detectors and reducing the cost of human annotation for new application environments. The proposal consists of several key research aspects: the principles and technologies of data association with multiple uncertain information sources, active learning with multiple information sources, transfer learning on weak annotated data, et al. We further propose following algorithms to handle above problem, i.e., transfer leaning based on multiple instance learning, incorporating co-training into active annotation, classification based on graphical model with multiple uncertain sources. If the proposal is implemented, we expect that the proposal would improve the development of smart visual surveillance, and provide the key algorithms and technologies for video surveillance.
英文关键词: smart visual surveillance;detector adaption;weak annotation;feature representation;data association