多摄像机协同的室内实时地图构建方法

An indoor real-time map construction method based on collaboration of multiple cameras

针对多摄像机时空信息提取精度低、融合效果差、与GIS集成困难的问题,提出了一种多摄像机构建室内实时地图的新方法。以高精度室内地图为基础,结合室内地理场景中的先验知识,计算各摄像机与室内地理空间的高精度单应矩阵;通过深度学习的方法,分别构建了人员头部前视和后视目标检测模型,实现了人员头部的高精度检测,并映射至地理空间;对映射后重叠区域中的目标,根据其地理方位、空间关系进行优选,实现了重叠区唯一目标的选择,最终和非重叠区目标共同构建出室内实时地图。基于办公室、教室多摄像机视频与GIS协同的实验结果表明:所提算法不仅实现了多摄像机下人员头部信息的高精度检测及地理映射,还能够准确地对重叠区域下的目标进行优选,研究结果可应用于办公场所人员监管、商业网点客流分析、交通客流统计等领域。

Aiming at the problems of low accuracy of multi-camera spatio-temporal information extraction,poor fusion effect,and difficulty in integration with GIS,a new method for building indoor real-time map with multi-cameras is proposed.Firstly,based on the high-precision map and the prior knowledge in the indoor geographic scene,the homography matrixes between each camera and geographic space are calculated.Secondly,through the deep learning method,the front-view and back-view target detection models of the head are respectively constructed,then the heads of multi-cameras are accurately detected and mapped to the geographic space;Finally,after mapping,the targets in the overlapping area are optimized according to its geographic position and spatial relationship,which realizes the selection of the only target in the overlapping area.An indoor real-time map is built,which contains the targets in the surveillance area.Based on multi-camera video and GIS maps in an office and a classroom,experimental results show that the proposed algorithm not only achieves high-precision detection and geographic mapping of person head information under multiple cameras,but also can accurately optimize targets in overlapping areas.The results can be applied to some fields such as the office person supervision,the flow analysis of commercial network customers and traffic passengers.