面向智能视觉监控的大规模慢特征学习研究

项目名称： 面向智能视觉监控的大规模慢特征学习研究

项目编号： No.61473290

项目类型： 面上项目

立项/批准年度： 2015

项目学科： 其他

项目作者： 张彰

作者单位： 中国科学院自动化研究所

项目金额： 80万元

中文摘要： 有效的特征表达对于完成视觉监控中的智能识别任务非常重要。人工设计特征（如方向梯度直方图等）已在行人检测、运动跟踪等监控识别任务中取得较好效果，但依旧不能完全解决由于监控场景变化而带来的环境适应性问题。本课题将借鉴视觉表达学习最新研究成果，开展监控环境下的大规模视觉表达学习研究。首先，我们将利用一般的运动检测、跟踪技术，从长时监控环境中收集海量运动目标数据；之后，我们将基于慢特征分析，这一针对有序样本的不变性特征学习方法，并利用监控场景环境约束所提供的弱监督信息，自下而上地进行层级式表达学习，获取监控环境下运动目标的多层慢特征表达；最后，我们将针对视觉监控识别任务，对一般性的慢特征做进一步优化，以提升识别性能和效率。通过本课题的研究，我们将建立一整套面向视觉监控的大规模慢特征学习方法，该方法将对开发自适应监控环境变化的智能识别算法带来极大帮助，从而有力推动智能视觉监控的研究发展。

中文关键词： 特征学习；智能视觉监控；慢特征分析

英文摘要： Effective feature representation is very important for developing intelligent visual surveillance applications. Current hand-crafted features, e.g., Histogram of Oriented Gradients (HoG), has achieved superior performance in pedestrian detection and tracking. However, it is still a hard problem to adapt the features to the variances in surveillance scenes. In this project, we will focus on the scalable representation learning for intelligent visual surveillance. Firstly, we will collect a large scale moving objects dataset in long-term surveillance scenes by common motion detection and tracking methods. Then, based on slow feature analysis (SFA) which can learn invariant features from sequential data, we will develop a scalable bottom-to-up representation learning method which is guided by the environment constraints in surveillance scenes to obtain multi-layers slow features of moving objects in surveillance scenes. Finally, according to the different surveillance tasks, we will further optimize the generic slow features to promote the performance and efficiency of the recognition algorithms. Through the research in this project, we will develop a whole set of scalable slow feature learning methods for visual surveillance. These methods are highly valuable for designing the adaptive recognition algorithms to handle the variances in surveillance scenes, so that the great progress in intelligent visual surveillance can be achieved.

英文关键词： Feature Learning;Intelligent Visual Surveillance;Slow Feature Analysis