基于机器视觉的监控视觉数据智能分析

Abstract：Surveillance visual data is one of the most important data resources of public security management. Safe city monitoring standard in our country is to set up the camera in an average of 200 meters spacing, so as to achieve full video surveillance. However, the uses of surveillance video mostly stay in post-analysis, manual observation, costly manpower, lack of effective intelligence analysis, resulting in a huge waste of resources. The project will build a library of intelligent analysis algorithms, aiming at public safety monitoring visual data and based on machine vision, simulate human visual perception and thinking strategy to intelligently understand and analyze the scene content, focusing on human / vehicle behavior as well as abnormal and unexpected phenomena. Around the intelligence analysis tasks, this project focuses on three parts: the intelligent segmentation of different functional areas in the monitoring scene; real-time accurate detection, identification and tracking of people, vehicles and traffic signs; rapid detection of public events (vehicle violation, pedestrian behavior, dangerous events, etc.). In view of the diversity of monitoring scenarios and monitoring tasks, the project will propose a series of innovative intelligent analysis techniques to apply in practical application environment, which based on deep learning, and especially make breakthrough on some key difficulties such as spatial and temporal correlation, migration learning, weakly-and semi-supervised learning, multi-object association.

监控视觉数据是公共安全管理最重要的数据资源之一。我国城市监控标准是按照平均200米间距布设摄像机，实现全程视频监控。然而，现下对监控视觉数据的使用多数还停留在人工观察和事后分析上，缺乏有效的智能分析，耗费人力资源巨大，造成了巨大的资源浪费。本项目将搭建一个针对公共安全监控视觉数据、基于机器视觉的智能分析算法库，模拟人类的视觉感知和思维模式智能地理解和分析场景内容，重点关注人/车行为、异常及突发现象，并对其作出准确提示或预警。围绕智能分析任务，本项目主要研究三部分内容：监控场景中不同功能区域的智能分割；人、车、交通标识的实时准确的检测、识别和跟踪；公共区域事件快速检测(车辆违规，行人行为，危险事件等)。针对监控场景和监控任务的多样性，本项目将提出一系列适用于实际应用环境、基于深度学习的创新性智能分析技术，特别在时空关联、迁移学习、弱监督/半监督学习、多目标关联等关键难点上取得突破。

针对监控场景复杂多变和监控任务的多样性特点，本项目结合多尺度表征学习、语义上下文感知学习、跨域跨模态关联学习、时空关联融合学习、基于不完备数据的弱/半监督学习和先验融合学习等创新思想，研究了基于监控视频图像的特定目标检测、显著性目标检测及异常事件检测，基于监控视频图像的目标识别与行为识别和基于监控视频图像修复等三方面内容，实现了对复杂监控任务的智能分析。本项目的研究成果解决了公共监控场景中多个关键科学问题，并取得了国际领先的效果，将有效推动城市公共安全监控的智能化建设。项目期间，共发表期刊论文13篇（其中SCI 2区及以上11篇，ESI高被引1篇），会议论文7篇（其中CCF A类7篇），申请中国发明专利8项，授权中国发明专利6项。基于所研发的关键技术，搭建了一个针对公共安全监控视觉数据、基于机器视觉的智能分析算法库，该算法库已集成到智能交通和智能航运等多个领域的创新性平台，实现了广泛的应用，产生了良好的社会和经济效益。获2021年度广东省科技进步二等奖（排名第一）。