实际视频场景中的连续行为识别较传统的单个行为识别更加具有挑战，本项目研究大数据背景下面向混合标记样本的连续行为识别技术。研究基于深度网络的行为时空特征自动学习方法，以数据驱动的方式学习更适应特定应用的特征。研究基于概率图模型的连续行为时序建模方法，以建立描述行为内动态过程以及行为间相互转移关系的数学模型。研究混合标记样本下集成模型的构建和学习方法，建立深度网络和概率图模型相结合的集成模型，以“端到端”的方式对混合标记样本进行学习，将行为建模的时序信息融入特征学习中，实现行为特征学习和连续行为识别的统一。本项目将面向实际应用环境，建立理论框架，设计关键算法，提高识别准确率，推动行为识别与分析技术的实用化。

Continuous action recognition in real video scene is more challenging than traditional isolated action recognition. This project researches on mixed labeled samples oriented continuous action recognition under the background of big data. Studying the method of automatically learning action time-space feature based on deep network, to learn features more adapt to specific application in a data-driven way. Studying time sequence modeling method for continuous action based on probabilistic graphical model, to establish a mathematic model describing the dynamic process within each action and transitions between actions. Studying the construction and learning methods of integrating model. Establishing the integrating model combining deep network and probabilistic graphic model. The integrating model learns from mixed labeled samples in end-to-end fashion, fusing the time sequence information of action modeling to feature learning and achieving action feature learning and continuous action recognition in a unified way. This project will face real application environment, conduct theoretical analysis and design key algorithms to improve recognition accuracy rate and promote the practical application of action recognition and analysis.