基于多载体协同跟随的多视角三维场景下人群异常事件检测方法研究

Visual surveillance on the crowds in public places is the important means for preventing "crowd disaster", diminishing all sorts of breaking events, ensuring safety of people and their properties. The classical intelligent video surveillance system based on single camera with fixed-position, however, limited by the position and viewing angle, is incapable of making real-time surveillance to large span zone, not to mention the active and continuous tracking on the object under surveillance, which greatly reduce the effect and efficiency of visual surveillance. In this proposal, based on the technology of mobile robot and intelligent video analysis, anomaly detection in crowd is investigated by the cooperative following behaviors of multiple carriers in 3D real scenes under different viewing angle. By combining video information obtained from the cameras mounted on fixed places with those on multiple carriers including flying robot, land robot, etc., We can not only realize real-time surveillance to large span zone with full-coverage, but also continuously track the object of interest and detect crowd anomaly as well. Our research topics include the following. First of all, real-time 3D scene is reconstructed based on the GPS information of multiple carriers and the global feature maps. Secondly,active object tracing based on collabration of multiple carriers.Thirdly, Abnormal behaviors of individuals in global or local zone are detected under 3D scene. It's expected the results of this research can facilitate early warning and analysis on crowd anomalies, which is of great practical use in reinforcing management in cities as well as maintaining public security.

对公共场所的密集人群进行视频监控是预防"人群灾难"，减少各种突发性事件，保障人民群众生命财产安全的重要手段。传统的基于固定单摄像机的智能视频监控系统由于位置和视角的限制，难以实现对大跨度场景的实时监控，更无法主动跟随监控目标进行连续追踪，从而影响监控的效果和效率。我们结合移动机器人技术和智能视频分析技术，提出基于多载体协同跟随实现多视角三维场景下的人群异常事件检测，通过融合固定摄像机及搭载在多载体（飞行机器人、地面移动机器人）上的移动摄像机的视频信息，不仅能对大跨度场景进行全覆盖式的实时监控，还能持续地追踪感兴趣的目标并进行异常行为检测。本项目研究内容包括：基于多载体自身GPS信息及全局多特征图的三维场景实时重建；多载体协同的目标主动追随；三维场景下对全局、局部及个体异常行为进行检测。项目研究成果能为各类人群异常事件提供早期预警和分析手段，对加强城市管理和保障公共安全具有重要的实用价值。

对公共场所的密集人群进行视频监控，是预防“人群灾难”，减少各种突发性事件，保障人民群众生命财产安全的重要手段。传统的基于固定的单个摄像机的智能视频监控系统，由于受到位置和视角的限制，难以实现对大跨度场景的实时监控，更无法主动跟随监控目标进行连续追踪，从而影响了监控的效果和效率。针对这一问题，我们提出了基于多载体摄像机来协同实现多视角三维场景下的人群异常事件检测方法。首先，我们通过移动机器人的辅助对多摄像机进行了协同标定，建立了适用于人群行为分析的三维场景模型，并且通过对虚拟人群和实际拍摄的人群进行视觉算法分析，验证了模型的可靠性。其次，我们通过对四元数离散余弦变换特性进行分析，在多尺度分析的基础上，利用反四元数离散余弦变换操作和高斯模糊处理，获得了时空异常显著热力图；之后通过概率论方法把多载体的异常目标主动跟随转化为一个决策问题，实现多载体对多个显著性目标的主动追随。最后，在近距离下，提出了移动相机下基于ORB-SLAM的三维背景估计的运动目标检测算法，并可在社区巡逻机器人上实时运行跑动检测算法；在远距离下，运用基于空间邻域加权NW-FCM聚类算法对人群运动模型进行建模；然后进一步完善了全局轨迹，并在此基础上进行了基于时间时延的大场景下人群异常检测分析，实现了在三维场景下的远距离检测和近距离检测相结合的人群异常事件检测和跟踪。通过项目开展，发表相关论文19篇，申请发明专利1项，软件著作权1项，达到了预期研究目标，部分核心技术在社区巡逻机器人上进行了应用。