主被动视角联合的细粒度行为识别

项目名称： 主被动视角联合的细粒度行为识别

项目编号： No.61502301

项目类型： 青年科学基金项目

立项/批准年度： 2016

项目学科： 计算机科学学科

项目作者： 倪冰冰

作者单位： 上海交通大学

项目金额： 22万元

中文摘要： 细粒度行为识别技术在智能辅助居住领域中有着极为重要的应用。然而，目前行为识别技术的研究主要着眼于粗粒度行为识别，缺乏细粒度行为识别的理论框架和算法模型。本项目结合申请人在视频行为识别方面的前期研究工作以及机器学习理论的发展趋势，分别从特征提取、语义表示、模型构建等方面对细粒度行为识别领域中的几个重要问题进行重点研究并提出解决方法，主要包括为：(1) 提出多任务联合深度特征学习、中层交互语义学习等新方法，有效解决细粒度动作和物体联合识别的问题。(2)提出基于非参数化贝叶斯多核学习的数据融合理论框架，实现主被动视角联合特征提取，解决多视角数据融合的难题，提高细粒度行为识别效果。(3) 提出基于递归网络的多层次联合分析模型，解决多层次结构化行为语义联合识别难题。本项目的研究成果将有力推动细粒度行为识别的研究及应用。

中文关键词： 视频行为识别；细粒度行为识别；多粒度行为识别；主被动视角相机联合行为识别；视觉特征学习

英文摘要： Video based fine-grained activity recognition and analysis has promising application in intelligent assisted living. However, state-of-the-art research work in activity recognition still focuses on coarse-grained activity, and novel theory framework, algorithm infrastructure, and recognition model for fine-grained activity recognition are still unexplored. In this project, we will investigate several key problems in fine-grained action recognition, which include (1) We propose to address joint motion-object visual feature extraction and middle level semantic learning problem in fine-grained activity recognition by introducing a novel multi-task joint deep network model; (2) We propose novel methods based on transfer learning and nonparametric Bayesian multiple kernel learning framework to achieve joint active-passive camera coordinated action recognition; (3) We plan to address the problem of multi-granularity joint analysis to achieve structural and hierarchical understanding, modeling and recognition of daily activity. The research outputs of this project will greatly encourage the research and application of activity recognition techniques in assisted living. Besides, this project can also provide novel theories and fundamental algorithmic frameworks for smart surveillance, remote medicine and video retrieval.

英文关键词： video based activity recognition;fine-grained activity recognition;multiple granularity analysis;active-passive camera joint analysis;visual feature learning