基于上下文学习的层次化行为分析与理解

Vision based activity analysis and understanding is an important way to implement the intelligent management of the massive video data. Human activity within the complex scenario often shows hierarchical structure, it often evolves from low to high levels and from simple to complex. In this proposal, we aim to study activity analysis and understanding problem and its hierarchy character based on context learning. We first study multi-level activity feature representation problem by fusing the spatial-temporal information to overcome deficiency of traditional methods. Through analysis from different levels of related information such as features, component, action, individual activity, group activity and the interactions among them, we explore the context learning method for the activity analysis problem from individual to groups. Then by introducing hierarchical context graph models, we mine the latent context structure relationship and information implied in group activities. By combing multi-task learning and deep learning within the same framework, we study the approach of joint learning for these hierarchical activities. Based on the activity analysis results, we also study the diversified semantic activities understanding problem existing in the complex scenarios, and establish a sparse topic model with one class constraint for the activity monitoring and safety alerting.

基于视觉的复杂场景中的行为分析和理解是实现海量视频数据的智能化管理的重要手段。针对视觉场景的复杂性，以及人体行为语义通常呈现出由底层到高层、由简单到复杂的层次化的特点，本项目旨在研究基于上下文学习的行为分析与理解方法及其层次化的演进过程。本项目首先研究融合时间-空间信息的多层次行为特征表示，克服传统方法在视觉行为特征表达方面的不足；进而从特征、部件、动作、个体行为、群组行为和复杂行为等不同层次的关联和交互信息入手，探讨行为上下文关联的学习方法；进一步引入层次化上下文图模型结构，挖掘群组行为中存在的隐含关系，实现由个体行为分析向群组行为分析的过渡；同时结合多任务学习与深度学习，研究多层次行为的联合学习和优化方法；最后基于场景行为分析的结果，探讨研究复杂场景中存在的多样化行为语义理解，建立基于一类约束的稀疏主题模型，为复杂场景中行为监测和安全预警服务。

随着视频监控系统覆盖率的不断提高以及存储技术的飞速发展，视频监控数据量正以惊人的速度增长。基于视觉的行为分析和理解能够自动对视频中的运动目标进行检测、分类和识别，是实现海量视频数据的智能化管理的重要手段。本项目研究基于上下文学习的行为分析与理解方法，研究基于视觉识别技术的多层次视觉特征表达方法，克服传统方法在视觉行为特征表达方面的不足；分析从视觉特征、目标检测和行为等不同层次的特征表示和语义信息提取方法，探讨研究复杂场景中存在的多样化行为语义，建立面向海量数据、低端设备的目标语义提取方法，为行为监测和安全预警服务。本项目的研究将建立对抗学习目标检测、跟踪、提取、分析的一整套方法，从而为智能视频分析的实际应用打下基础，同时推动学术研究的发展。