面向类人概念学习的视频事件时空分析的几何代数方法研究

Human-level concept learning of video events has emerged as one of the key technologies in the forefront of machine learning, owing to its effective capability of solving the long-standing deep learning problem – detection and recognition of small-sample video events, a critical need and key technique for the development of our society and economy. .Spatio-temporal analysis lays the basis and serves as one of the key techniques for human-level concept learning of video events. However, traditional spatio-temporal analysis techniques cannot meet the intense needs of human-level concept learning, due to limitations of the widely adopted Euclidean geometric framework. In particular, the Euclidean geometric framework is insufficient to uniformly model, describe and calculate video events of diverse dimensions and types. Moreover, it is difficult to perform efficient and complete analysis of the complex spatio-temporal relationship for video events. .As a remedy, this project carries out innovative research on spatio-temporal analysis of video events from the fundamental mathematical framework. Specifically, we will develop novel spatio-temporal analysis theory and methods for human-level concept learning based on the geometric algebra, which resolves the deficiency of the traditional approaches. Moreover, based upon our existing works in this field, we will further realize the human-level concept learning system of video events embracing compositionality, causality and learning to learn. This project will effectively promote the development of human-level concept learning and machine learning of small-sample video events.

视频事件类人概念学习是当前机器学习最前沿的技术之一，能为深度学习尚未解决的小样本视频事件的检测与识别提供有效的方法，是社会经济发展中迫切需要的技术。.视频事件时空分析方法是视频事件类人概念学习的基础和关键技术之一，但是，传统的视频时空分析技术难以支撑视频事件类人概念学习，究其原因是它所采用的欧式几何框架难以对含有不同类型、不同维度数据的视频事件进行统一地建模、描述和计算，难以对复杂的视频事件时空关系进行高效完备地分析。.因此，本项目从底层数学框架对视频事件时空分析理论和方法展开源头创新研究，发展一套基于几何代数的面向视频事件类人概念学习的视频事件时空分析理论与方法，从底层数学框架上弥补现有视频时空分析技术的不足。结合我们现有的视频事件类人概念的工作，实现具有组合性、因果性、自学习性的视频事件类人概念学习系统。本项目将能有效促进类人概念学习和视频事件小样本机器学习的发展。

本项目对基于几何代数方法的面向类人概念学习的时空分析展开研究，完成了研究计划。首先研究了视频的几何代数统一模型，对不同类型、不同维度的对象进行几何代数建模、描述和计算。接着研究了几何代数空间中视频事件或复杂行为的多层次多粒度结构组成，提出了视频群体事件的多层次多维表示、基于时空兴趣点时空分布的行为时空多粒度分割、面向交通视频的交通轨迹多层次多粒度表示等；然后研究了面向视频事件类人概念学习的视频事件时空关系描述、表示及推理，提出了基于几何代数的视频群体事件时空关系的多维统一描述和度量、基于几何代数旋转算子的多人交互的时空关系描述及度量、多人交互时空关系描述及度量、交通事件复合定性时空推理方法等。最后在相关应用中进行验证。.发表论文20篇；发明专利11个，软件著作权2个；组织或参加学术会议10人次，培养博士生4名，硕士生8名。相关技术应用于康佳集团电视内容管理系统中，获广东省科技进步奖。