摘要
为获取全方位的人流量信息,实现对智慧建筑内资源的合理调配和人员的规范化管理,提出一套融合深度信息的双向人流量统计方法。该方法提出一种结合目标位置相似度与深度相似度的目标匹配策略,用于提升轨迹关联的可靠性和准确性;为减少目标频繁遮挡导致轨迹匹配错误以及目标ID频繁切换的情况,提出利用目标的轨迹上下文深度差判断人员遮挡状态,并优化轨迹处理;提出利用三维质心坐标判断人员横、纵运动方向,处理相机视角变化和不规则运动,减少方向判断错误。实验结果表明,所提方法在人员快速流动、分组聚集以及排队移动等场景下相比前人工作具有更低的漏检率和误检率,部分场景双向人流量统计的F1值达到100%,并且在配备有NVIDIA GTX1080Ti的主机上能保持20~35 F/s的实时速度,具有较高的实用价值。
In order to obtain two-way pedestrian flow information and realize the rational allocation of resources and standardized management of personnel in the smart building,a set of two-way pedestrian flow statistical method integrating depth information is proposed. The method proposes a target matching strategy combining the similarity of target position and the similarity of target’s depth to improve the reliability and accuracy of trajectory correlation. In order to reduce the trajectory matching error caused by the frequent occlusion of the target and the frequent switching of the target ID,it is proposed to use the depth difference of the trajectory context to judge the occlusion state of the target and optimize the trajectory processing. It is proposed to use the three-dimensional centroid coordinates to judge the horizontal and vertical motion direction of the person,which can deal with the camera angle change and irregular motion,and reduce the error of direction judgment. Experimental results show that the proposed method in the fast-moving, group gathered and queuing mobile scenarios compared to previous work has lower miss rate and false detection rate. For part of the scene,F1 value of twoway pedestrain flow statistical method reaches 100%,and the hosts equipped with NVIDIA GTX1080 Ti can keep 20 ~ 35 F/s real-time speed,which has high practical value.
作者
张文利
王宁
郭向
杨堃
马超
朱清宇
ZHANG Wen-li;WANG Ning;GUO Xiang;YANG Kun;MA Chao;ZHU Qing-yu(Faculty of Information Technology,Beijing University of Technology,Beijing 100124,China;China Construction Technology Group Co.,Ltd.,Beijing 100070,China)
出处
《测控技术》
2021年第5期52-61,共10页
Measurement & Control Technology
基金
“十三五”国家重点研发计划项目(2019YFE0100300)。
关键词
RGB-D
双向人流量统计
深度差
三维质心坐标
RGB-D
two-way pedestrian flow statistics
depth difference
three-dimensional centroid coordinates
