项目名称: 注意选择引导的人体运动分析和识别
项目编号: No.61472387
项目类型: 面上项目
立项/批准年度: 2015
项目学科: 计算机科学学科
项目作者: 卿来云
作者单位: 中国科学院大学
项目金额: 80万元
中文摘要: 人体运动分析是计算机视觉领域的研究热点之一,但由于人体运动的复杂性和视频数据标注困难的影响,人体运动分析仍是一个很有挑战的问题。另一方面,生物视觉系统因其良好的注意选择和采用多通道层级表示及相互协同,在运动分析上表现了显著的优越性。本项目拟借鉴生物视觉系统的信息加工机制,研究:(1) 快速轮廓特征和运动特征提取及注意选择建,提高视频数据的处理速度;(2) 无监督或视点弱标注的学习,及其在运动物体部件检测中的应用,为视频数据的处理提供标注费用和学习效率的一个折中;(3)运动部件的时空关系学习,及其在人体运动识别和定位中的应用。项目预期发表一些高水平文章,并开发一个人体运动识别和定位原型系统,可用于智能视频监控或人机交互。
中文关键词: 注意选择;运动分析;层级模型;弱标注
英文摘要: Human activity recognition is one of the hot topics in computer vision. Though many progress has been made in the last few years, it still a challenge problem because of the complexity of human behavior and the lack of enough labeled training data. Inspired by the select attention scheme in human vision system, we are going to develop a hierarchical visual representation used in human activity recogntion. More specifically, we aim to (1) modeling the selection attention based on objectness to locate the potential object quickly; (2) propose novel learning methods for detecting the discriminative parts without or with only the fixations; (3) propose novel learning method for modeling the spatial layouts and timing variations of moving object parts; (4) develop human activity recogntion systems. This project aims to output high quality research papers as well as effective real-world application systems.
英文关键词: Attention;Motion Analysis;Hierarchal Representaiton;Weak labeded Learning;fixation