云环境下基于图形化预定义和参数化方法的视频对象行为分析关键技术研究

Intelligent behavior analysis is the public research base of national priority research fields, including the locking of the military targets in national defense, the judgment of the aircraft flight status, the real-time analysis of astronauts' extravehicular behavior and public safety monitoring, and so on, which is the key to the success of the above studies. This research focuses on the graphical predefinition of video objects and their expected behavior, fast parametric skeleton modeling, parameter modeling of the dynamic behavior process, the creation of the space of behavior analysis based on prior knowledge , Semantic Transformation, the estimation of the head pose and gaze direction, cloud services and other key technologies. It proposes a series of new methods involving the parametric judgment of behavioral status and the analysis of behavioral process based on the multi-objective skeleton modeling and the analysis of local behavior characteristics based on the estimation of head pose and gaze direction. Then it will be applied to any object under a custom environment to build a shared public center providing services of video retrieval and behavior analysis in cloud environments by graphical predefinition. This research makes the intelligent analysis applicable, universal, public, ministrant, and lays the Theoretical foundation for the infrastructure construction and global application of multi-field intelligent video behavior analysis in our country.

智能行为分析是国防军事目标锁定、飞行器飞行状态判定、宇航员出舱行为实时分析及公共安全监控等国家重点研究领域的公共研究基础，是上述研究成败的重要因素之一。本课题重点研究视频对象及其预期行为的图形化预定义、快速参数化骨架建模、动态行为过程的参数模型建立、基于先验知识的行为分析空间建立及语义转换、头部姿态及视线方向估计、云服务等关键技术，提出基于多目标骨架建模的参数化行为状态判定及行为过程分析、基于头部姿态和视线方向估计的局部行为特征分析等一系列核心新方法，并通过图形化预定义的方式，将其应用于自定义环境下的任意对象，构建云环境下共享式视频检索和行为分析公共服务中心。该课题的研究能够实现智能分析的应用化、普及化、公共化、服务化，为我国多领域智能性视频行为分析的基础设施建设及全球性应用推广奠定一定的理论基础和技术支撑。

智能行为分析是国防军事目标锁定、飞行器飞行状态判定、宇航员出舱行为实时分析及公共安全监控等国家重点研究领域的公共研究基础。本项目对智能行为分析中骨架建模、参数化建模及行为状态判定等问题进行了重点研究。首先，对传统光流法计算量大、实时性差的问题进行了改进，提高了运动目标检测与跟踪效率，开发完成了视频环境下运动目标检测跟踪系统。其次，对提取的运动人体轮廓进行了优化，将轮廓分为水平、垂直和曲线三个区域，并在不同区域采用相应细化算法得到完整的骨架。最后，基于骨架建模和参数建模的人体行为趋势分析,提出了改进的人体骨架关键点定位算法，并基于关键点建立了人体参数模型，将人体行为趋势分析问题转化为参数求解问题，大大减少了人体行为分析的计算量和存储量，提高了运行速度，开发完成了行为分析状态判定软件。