【作者】 张华；

【导师】 刘宏立； 黎伟；

【作者基本信息】 湖南大学 ， 电子与通信工程（专业学位）， 2015， 硕士

【摘要】 人头检测与统计系统是智能监控中重要的研究方向,有着广泛的应用价值。在地铁站和火车站等重要公共场所,人数统计为控制客流提供精确的客流数据管理;在公交车和电梯口等区域,客流量统计对公共安全防范和交通配置都有着重要的参考价值;对于大型超市和商场等企业,在人数统计的基础上进行相关商业分析和数据挖掘可以辅助企业指导市场决策。因此,人数统计带来的社会意义和市场应用价值使得其成为当前国内外智能视频监控领域的研究热点。本文基于当前先进的立体视觉技术来进行人头检测与统计方法研究,主要研究了基于双目立体视觉的被动立体视觉技术在人头检测与统计的设计与实现,然后针对被动立体视觉的局限性结合当前先进的主动立体视觉技术,设计了人头检测与统计方法在深度摄像头上的实现。在双目立体视觉人头检测与统计方法研究中,首先进行双目摄像头的标定工作,求出摄像头的内外参数和校正矩阵。然后利用校正后的图像进行立体匹配获取视差图,使用了一种跨尺度变窗口代价聚合的快速立体匹配方法。最后,利用俯视深度图并结合人头距离摄像头距离最近的特点检测出人头。在整个过程中,提出并实现了一种低复杂度高精度的立体匹配算法,然后改进了一种基于深度图的人头检测和统计方法。实验结果表明,基于双目立体视觉的人头检测与统计方法成本低,不仅检测精度很高,还能适应各种复杂的环境变化,从而鲁棒性较强。在基于深度摄像头的主动立体视觉方法中,提出一种适用于深度图的模拟降水分水岭算法(Depth map based Rainfalling Watershed Segmentation,D-RWS)。首先,修复深度图的空洞,并用高斯混合模型的背景建模方法进行前景图像的提取;然后,利用D-RWS算法在深度图中分割出感兴趣的行人头部区域(Region Of Interest,ROI);最后,基于质心欧式距离最短法的原则关联各图像帧中的相同目标,根据跟踪的轨迹来统计。通过仿真和软件实现此算法,实际场景测试表明主动立体视觉的人头检测和统计方法达到了40帧每秒,实用性很强,检测精度和效率高。最后总结了主动和被动两种基于立体视觉的方法人头检测和统计方法,两方法均可解决传统方法因行人遮挡、光线突变导致的检测准确率低的难题。双目的方法鲁棒性强,但是实时性能略微逊色。深度传感器的方法由于采用结构光技术也导致其视野范围以及有效距离都较小,受太阳光中的红外成分影响而不能适应室外环境。在经济因素、室内外应用场合和实时性能要求不同的应用场景下,可根据实际需要选择其中一种合适的解决方案。更多还原

【Abstract】 Head detection and counting system is an important research direction of intelligent control, it has extensive application value. In public places such as subway and train stations, the number of passenger flow statistics provide data management for accurate control of passenger flow; In the bus and the elevator, traffic statistics have important reference value for public security and traffic allocation; For large supermarkets and shopping malls and other enterprises, the use of the number of statistical data related to business analysis and data mining can help enterprise to guide the market decision. Therefore, the number of statistics about social significance and market value makes it become the research hotspot in the field of intelligent video surveillance.Technique based on the stereo vision is advanced to the head of the detection and counting method, this paper mainly focuses on the research of binocular stereo vision of passive stereo vision technology in the design and implementation of a head detection and counting based on, and then aiming at the limitation of passive stereo vision with the advanced active stereo vision technology, design and realization of the detection and counting methods in depth on the camera head.In the research of detecting and counting method of binocular stereo vision in human head, first of all, the calibration of binocular camera internal and external parameters of the camera are obtained.Then, the disparity map is get by using a fast stereo cross scale dynamic support windows cost aggregation matching method. Finally, combine the vertical depth map and the characteristics of the camera head distance distance recently detected the man’s head. In the whole process, this paper designs a new stereo matching algori thm, and improved the disparity map head detection and counting algorithm based on. The experimental results show that the head detection and counting method based on binocular stereo vision of low cost, not only high accuracy, but also can adapt to the co mplex environment changes, and the strong robustness.Based on the active stereo vision method in depth camera, a novel method D-RWS(Depth map based Rainfalling Watershed Segmentation)is proposed. Depth map is inpainted and foreground is extracted with the help of mixture of Gaussian background model. D-RWS algorithm is used to segment head area as Region Of Interest(ROI). People are tracked and counted by analyzing trajectories, which associated by minimal Euclidean distance between the centers. Experiment al results show that proposed people counting system can, on average, count people with an accuracy of 98% and operate at approximately 25 milliseconds per frame(40 f/s).The accuracy and real-time performance fully meet the requirements of practical application.Finally, this paper summarizes the active stereo vision and passive stereo vision method to detect and count two heads. It Solved the traditional method for pedestrian block and light mutations can lead to the problem of low accuracy. Binocular metho d is robust, but real-time performance slightly less. Methods the depth sensor with structured light technology has also led to the range of its vision and the effective distance is smaller, affected by the infrared component in the sunlight can not adapt to the outdoor environment. In economic factors, indoor and outdoor applications and real-time performance requirements of different application scenarios,we can according to the actual need to select one of the appropriate solution.更多还原