视频GIS中行人三维精确定位方法研究

3D precise localization and analysis of pedestrians in an integrated system of real-time video and 3D GIS will have important research significance and application value in social security and management. Considering the problems of inaccurate positioning and jumping trajectory of pedestrians which result from shielding, shadows and light changes in video 3D GIS, combined with the TensorFlow platform, a GPU-accelerated MDNet-based pedestrian tracking method is proposed to detect, track and extract information of moving pedestrians, and a pedestrian image database is established to provide basic data for subsequent researches. A DCGAN-based image completion method is proposed to restore incomplete pedestrian images and provides complete foreground images for 3D precise localization of pedestrians. A 3D precise localization and motion simulation method for multi-pedestrians is implemented to establish the associations between pedestrians and 3D virtual scene, and to realize 3D visualization and motion simulation of multi-pedestrians in Video GIS by displaying pedestrian images, 3D trajectories and annotations. The accuracy of the 3D precise localization method is evaluated by comparing the 3D trajectories created by this method with the reference trajectories drawn by a pedestrian trajectory acquisition method based on 3D laser point cloud data and 3D model constraints. This research can improve the accuracy of 3D localization of pedestrians in Video GISs, and provide methodological and technical support for applications such as public safety, intelligent security, indoor positioning and navigation, and building/park management.

将实时视频数据与3D GIS相融合、并对运动行人进行三维精确定位与分析，在社会安全与管理方面具有重要的研究意义与应用价值。针对视频3D GIS中因遮挡、阴影、光线变化等因素导致的行人三维定位不准确、运动轨迹跳跃等问题，本项目结合TensorFlow深度学习平台，研究GPU加速的MDNet行人跟踪方法，对运动行人进行检测、跟踪与信息提取，并建立行人图像库；提出基于DCGAN的轮廓不完整行人图像补全方法，对不完整的行人图像进行补全，为行人三维定位提供完整的前景图像信息；研究多行人三维精确定位与运动模拟方法，建立多个行人与三维虚拟场景的关联，实现行人图像、三维轨迹、属性标注等在视频GIS中的三维可视化与动态模拟，并对三维定位精度进行评估与分析。本研究能够提高行人在视频GIS中的三维定位精度，为公共安全、智能安防、室内定位导航、楼宇/园区管理等应用提供研究基础与技术保障。

随着公共安全、楼宇/园区/企业厂区安全监控与管理对人员定位要求的提高，将视频数据与3D GIS相融合，并对运动行人进行三维定位、识别、跟踪与分析，对社会安全与管理具有重要意义。本项目围绕视频GIS中运动行人三维精确定位这一关键问题，结合深度学习平台和视频GIS平台，研究融合YOLOv7与SiamMask的行人跟踪方法，对运动行人进行检测、跟踪与信息提取，并建立行人图像库；研究基于AOT-GAN的轮廓不完整行人图像补全方法，对不完整的行人图像进行修复和补全，为行人三维定位提供完整的前景图像信息；研究多行人三维精确定位与运动模拟方法，建立多个行人与三维虚拟场景的关联，实现行人图像、三维轨迹、属性标注等在视频GIS中的三维可视化与动态模拟，并对三维定位精度进行评估与分析。实验结果表明，本项目算法解决了行人三维定位不精确、运动轨迹跳跃等问题，视频行人三维定位精度达到10-15cm。自主研发了“实时视频行人三维定位与运动模拟系统”，有效地将计算机视觉相关技术与3D GIS相融合，增强了三维场景的实时性和动态性，为视频行人三维定位算法的实施与精度评估提供了工作平台与验证平台。本项目积累了大量关键数据，包括实验场地实测数据、建筑物CAD数据、建筑物三维模型、倾斜摄影数据、部分实验场地的三维激光点云数据、监控视频数据、行人图像库等，能够为视频行人三维定位的研究拓展与应用落地提供宝贵的数据支撑。通过本项目的算法研究与系统研发，实现了多行人在三维场景中的精确定位与运动模拟，发展了视频GIS中运动行人三维精确定位方法，在公共安全、智能安防、室内定位导航、楼宇/园区管理等方面具有广阔的发展前景和应用价值。