冬奥村全景视频数据
收藏资源简介：
大尺度建图通常采用运动恢复结构技术，而传统的运动恢复结构只利用视觉信息，在对称结构、重复纹理、弱纹理的场景鲁棒性不足，并且需要频繁地进行大量图像位姿和三维点坐标的全局优化，效率不高。本项目创新地提出了多源数据融合建图方案，把惯性测量单元的局部相对位姿，GPS 的全局位置作为先验应用于视觉特征匹配，同时利用全景图像大视场的特点，减轻重复纹理的影响，形成鲁棒的数据关联，可以在困难场景下正确解算设备位姿和场景结构； 在优化方面，通过数据分块将全局捆绑调整问题转化为块内的局部捆绑调整和整体的全局位姿图优化相结合的问题，提升建图的效率。本数据集是使用Insta360 ONE X2运动相机在冬奥运村范围内可行走道路行走进行采集。把惯性测量单元的局部相对位姿，GPS的全局位置作为先验应用于视觉特征匹配，同时利用全景图像大视场的特点，减轻重复纹理的影响，形成鲁棒的数据关联，可以在困难场景下正确解算设备位姿和场景结构。
Large-scale mapping typically relies on Structure from Motion (SfM) techniques. However, traditional SfM only leverages visual information, exhibiting insufficient robustness in scenarios with symmetric structures, repetitive textures, and low-texture regions. Furthermore, it requires frequent global optimization of numerous image poses and 3D point coordinates, leading to low mapping efficiency. This project proposes an innovative multi-source data fusion mapping framework. It applies the local relative poses from the Inertial Measurement Unit (IMU) and the global positions from GPS as priors to visual feature matching. Meanwhile, leveraging the wide field-of-view of panoramic images, it mitigates the impact of repetitive textures to form robust data associations, enabling accurate estimation of device poses and scene structures in challenging scenarios. For optimization, this work transforms the global bundle adjustment problem into a combination of intra-block local bundle adjustment and overall global pose graph optimization via data partitioning, thereby improving the efficiency of mapping. This dataset was collected by walking on passable roads within the Winter Olympic Village using an Insta360 ONE X2 action camera. The dataset applies the local relative poses from IMU and global positions from GPS as priors to visual feature matching, leverages the wide field-of-view of panoramic images to alleviate the impact of repetitive textures, forms robust data associations, and enables accurate estimation of device poses and scene structures in challenging scenarios.
