异常行为识别是数字视频内容安全领域的重要研究方向，目前普遍受到：基于画面内容的行为描述算法局限性、基于标注数据和监督训练的建模算法局限性、数据隐私性保护局限性三个方面的限制，其发展正逐渐面临来自真实开放场景应用的挑战。本项目从开放环境下的行为安全检测需求角度出发，利用骨骼数据冗余小、隐私性高等优点，探索骨骼数据表征和图深度学习理论结合下的异常行为识别技术，重点解决骨骼序列的图结构化表征，多视角骨骼姿态生成，弱标注行为样本训练三个核心问题，分别提出基于图卷积网络的骨骼序列隐层表征算法、基于图循环网络的多视角骨骼序列生成算法、基于图自编码网络的无监督数据学习算法等研究内容。不同于根据姿态、运动等人的先验知识识别异常行为的传统思路，本项目将人的骨骼结构转化为图结构，赋予模型挖掘图结构数据深层关联的能力，有望突破现有识别方法框架的局限，引领形成新型的开放场景异常行为识别框架。

Abnormal behavior recognition is an important research area in the field of digital video content security. At present, it is generally limited by three aspects: the limitation of behavior description algorithm based on video content, the limitation of modeling algorithm based on annotation and supervised learning, and the limitation of data privacy protection. The development is gradually facing the challenge from the real open scene application. From the point of view of the needs of behavior security detection in open environment, the applicant explores the abnormal behavior recognition technology based on the combination of skeleton data representation and graph learning theory. Taking advantage of small skeleton data redundancy and high privacy, we focus on three main problems: the representation of skeleton sequence structure, the generation of cross view skeleton posture, and the training of weak annotated behavior samples. The research contents include semi-supervised abnormal behavior recognition algorithm based on graph convolution network, cross perspective abnormal behavior recognition algorithm based on graph recurrent network, unsupervised abnormal behavior recognition algorithm based on graph auto-encoder network, etc. Different from the traditional idea of identifying abnormal behaviors based on the prior knowledge of posture, motion, etc., this project transforms human skeleton structure into graph structure, endows the model with the ability of mining the deep association of graph structure data, which is expected to break through the limitations of the existing recognition method framework and lead to the formation of a new open scene abnormal behavior recognition framework.