基于深度区域特征和概念共现网络的人体异常行为识别

Human abnormal behavior recognition guarantees the rapid development of intelligent video surveillance technology, and it is one of difficulties and hotspots in computer vision. We focus on abnormal behaviors of human body such as crowd gathering, hand-held weapons, fast running, etc., which happened commonly in crowd places. The project dedicates to study human abnormal behavior recognition based on deep network construction and network analysis methods of deep walk. The specific research contents are as follows: (1) Based on the multi-task R-CNN (Region-based Convolutional Neural Networks), the project aims to study region feature extraction, semantic concept extraction and anomaly detection methods, which combined with self-attention mechanism to fully exploit inter-class differences and intra-class compactness. (2) Based on the complex network modeling method of word activation forces (WAFs), this project studies the methods for image semantic concept network construction and optimization. So as to explore the low-dimensional structure of high-dimensional data. (3) Corresponding to constructed semantic concept network, based on deep walking, this project studies the methods for semantic concept co-occurrence patterns mining and scene concept extracting, and achieves human abnormal behavior recognition after feature fusion. This research project and its achievements are of great significance for enriching the research of abnormal behavior recognition on basis of image understanding, and promoting the development of public safety monitoring system for “smart city”.

人体异常行为识别是智能视频监控技术高速发展的保障，是计算机视觉领域的主要研究任务之一。针对人群聚集场所中普遍存在的人体异常行为如人群聚集、手握凶器、快速奔跑等，本项目拟基于区域特征和词激活力实现深度网络构建，并结合深度游走的网络分析方法，开展公共视频中的人体异常行为识别方法研究。具体研究内容包括：（1）基于多任务区域卷积神经网络，研究区域特征提取、语义概念提取和异常区域检测方法，融合自注意力机制充分挖掘类间差异性和类内紧致性；（2）基于词激活力的复杂网络构建方法，研究语义概念网络构建和优化方法，为挖掘高维数据的低维结构提供基础；（3）在构建语义概念网络的基础上，基于深度游走，研究语义概念共现模式挖掘和场景语义概念提取方法，通过特征融合，实现人体异常行为识别。该项目研究及其成果对于丰富基于图像理解的人体异常行为识别的研究内涵，推动智慧城市中公共安全监控系统的发展具有重要意义。

人体异常行为识别是智能视频监控技术高速发展的保障，是计算机视觉领域的主要研究任务之一。针对人群聚集场所中普遍存在的人体异常行为如人群聚集、手握凶器、快速奔跑等，本项目基于深度区域特征提取实现异常检测网络构建，开展公共视频中的人体异常行为识别方法研究。具体研究内容包括：(1)研究深度区域特征提取和异常检测方法，融合注意力机制充分挖掘类间差异性和类内紧致性；(2)研究共现模式挖掘和语义特征提取方法，实现特征融合和增强，为挖掘高维数据的低维结构提供基础；(3)研究复杂深度网络构建和优化方法，实现高准确率的人体异常行为识别。该项目研究及其成果对于丰富基于图像理解的人体异常行为识别的研究内涵，推动智慧城市中公共安全监控系统的发展具有重要意义。