面向群体行为识别的渐进层级性分析模型研究

Multi-level in-depth analysis of crowd activity recognition is an important issue to be solved in the field of activity recognition. The current research on activity recognition is not limited to recognize the categories of the activities, it pays more attention to the details of the activities. The existing research expands the depth and breadth of activity recognition, and its method system is far from perfect. Based on the previous research of machine learning, especially deep neural network, this project proposes a progressive hierarchical analysis model for activity recognition. The study mainly focuses on the research of a temporal consistency detection model for multi-person objects based on transfer learning, a research which integrating spatio-temporal feature learning for recognizing individual actions in the crowd with unconstrained action duration, and a research on crowd activity recognition method based on individual action, context of interaction and scene. These studies have made a series of innovations, without information annotation, multi-person in the crowd are detected through the modulating network based on transfer learning, multi-person detection and tracking tasks are completed through the temporal consistency detection approach, the individual actions in the crowd are recognized with unconstrained action duration, and integrating the individual action, interaction context and scene context, the crowd activity is steady and effectively recognized. The research results of the proposed model can not only enrich the methods and theories of machine learning, but also produce important social and economic value.

群体行为的多层次深度分析是行为识别领域亟待解决的重要问题。目前行为识别的研究不在局限于行为类别的识别，而是更加的关注行为群体的活动细节。现有研究拓展了行为识别的深度与广度，其方法体系远未完善。在以往机器学习特别是深度神经网络研究的基础上，本课题提出了一个面向群体行为识别的渐进层级性分析模型，主要进行了基于迁移学习多人体目标的时序一致性检测模型研究；融合时空特征学习时长无约束的群体中个体行为识别方法的研究；结合个体行为与交互场景上下文的群体行为识别方法研究。这些研究做出了一系列创新，无需信息标注，通过可调控网络的迁移学习实现了行为群体中的多人体检测；通过时序一致性检测完成了多人体目标的检测与跟踪任务；完成了时长无约束的行为群体中的个体行为识别；并融合人体行为，交互、场景上下文信息，对群体行为进行了稳定有效的识别。所提模型取得的成果不仅能丰富机器学习的方法和理论，还能产生重要的社会经济价值。

对群体行为进行多层次的深度分析是行为识别领域亟待解决的重要问题。伴随着深度神经网络的发展，循环神经网络RNN与LSTM展现了其在时序数据处理方面的优势，本研究基于深度学习网络架构，主要解决了以下三个问题：1，迁移人体检测器，实现多人体目标的时序一致性检测；利用时序长度不确定的训练样本，训练行为分类器；3，利用场景上下文与行为交互上下文信息提升群体行为的识别效果。相应的具体研究内容为：1，基于调控网络的迁移学习，实现了多人体目标的时序一致性检测；2，融合时空特征学习，实现了群体行为中时长无约束的个体行为识别；3，通过场景中个体行为类别、交互场景上下文信息的融合，实现了对群体行为稳定有效的识别。取得的重要结果和科学意义主要包括：基于迁移学习、深度神经网络架构，拓展了行为识别的深度与广度，设计了群体行为识别的层级性分析模型，实现了端到端的群体行为识别；在公用数据集上进行的大量实验表明，与现有方法相比，该模型在群体行为分析识别方面具有良好的效果；所设计的层级性的群体行为分析模型，能广泛的应用于视频监控、网络视频中的人体行为分析等。