面向智能视频监控的多目标检测与跟踪技术研究

With the deployment of more and more surveillance cameras, research on computer based intelligent video surveillance (IVS) techniques is becomeing more important. IVS is usually based on computer vision and pattern recognition theories and its goal is to analyze and response promtly to the video content. IVS can also provide useful statistical information, such as the number, the distribution, the intensity and the flow of the targets, to those who are interseted in it. However, existing methods are far more behind people's requrements in terms of accuracy and the real-time constraint. Especially in complex situations, such as the crowded traffic scenes, occlusion between targets and occlusion casued by scene static objects can seriously degrade the performance of object detection and tracking mehtods. Cosidering the disadvantages of existing object detection and tracking methods, this project aims at studying the following problems when occlusion exists prevalently: 1) fast multi-pedestrian detection based on 2D-3D reasoning; 2) effective multi-pedestrian tracking based on the information provided by the visible parts of the targets; 3) making use of temporal and spatial information simultaneously to estimate the scene occluders to compensate both the detection and tracking algorithms; 4) multiple targets detection and tracking when pedestrians, bicycles and vehicles exist in the same scene. Our final goal is to process real world survillance videos reliably in real-time.

随着监控摄像头安装的日益增多，计算机智能视频监控技术的研究变得越来越重要。它基于计算机视觉和模式识别等理论，以对视频内容做出快速的分析和响应为目标，并能提供有用的统计数据如目标数量，分布，密度及流量。但是现有技术在准确性和实时性方面远不能满足人们的要求。特别是在复杂情况下，例如比较拥挤的交通场景，目标之间的相互遮挡，以及场景物体对目标的遮挡，都会严重影响目标检测和跟踪算法的准确性。基于当前研究在目标检测和跟踪方面存在的不足，本项目着力研究在遮挡普遍存在的情况下，1）利用从二维到三维的推理来实现对多个行人的快速检测方法，2）最大化利用目标未遮挡部分提供的信息进行有效跟踪的方法，3）同时利用时间和空间信息对场景障碍物进行估计以实现对检测和跟踪补偿的方法，4）车辆和行人同时存在情况下的检测和跟踪。本项目将力求达到可靠地对实际的监控视频进行实时处理，为未来智能监控的实际应用提供理论依据和技术支持。

随着监控摄像头安装的日益增多，计算机智能视频监控技术的研究变得越来越重要。它基于计算机视觉和模式识别等理论，以对视频内容做出快速的分析和响应为目标，并能提供有用的统计数据如目标数量，分布，密度及流量。但是现有技术在准确性和实时性方面远不能满足人们的要求。特别是在复杂情况下，例如比较拥挤的交通场景，目标之间的相互遮挡，以及场景物体对目标的遮挡，都会严重影响目标检测和跟踪算法的准确性。基于当前研究在目标检测和跟踪方面存在的不足，本项目研究了在遮挡普遍存在的情况下，1）基于对可变性部位模型的分解和遮挡处理实现来实现多个行人的准确检测，2）利用尺度估计和遮挡推理实现多行人的检测，3）迭代的进行数据关联和检测结果更新来实现准确的多目标跟踪，4）同时利用时间和空间信息对场景障碍物进行估计以实现对检测和跟踪补偿的方法，5）利用卷积神经网络特征和相关滤波器实现行人和车辆的联合跟踪 。本项目将为未来智能监控的实际应用提供理论依据和技术支持。