HSC4D

我们提出了以人为中心的4D场景捕获 (HSC4D)，以准确有效地创建动态数字世界，其中包含大规模的室内外场景，多样化的人类运动以及人类与环境之间的丰富互动。仅使用车身安装的IMUs和激光雷达，HSC4D是无空间的，没有任何外部设备的约束，并且没有预先构建的地图。考虑到imu可以捕获人类姿势，但在长期使用中总是漂移，而LiDAR对于全球定位是稳定的，但对于本地位置和方向是粗糙的，HSC4D通过联合优化使两个传感器相互补充，并在长期捕获中取得了有希望的结果。还探索了人类与环境之间的关系，以使其相互作用更加现实。为了方便许多下游任务，如AR、VR、机器人、自动驾驶等，我们提出了一个包含三个具有精确动态人类的大场景 (1k-5k m2) 的数据集

We propose Human-Centered 4D Scene Capture (HSC4D) to accurately and efficiently create dynamic digital worlds that encompass large-scale indoor and outdoor scenarios, diverse human motions, and rich interactions between humans and their surrounding environments. Utilizing only body-mounted IMUs and LiDARs, HSC4D is free from external device constraints and does not require pre-built maps. Given that IMUs can capture human poses but inevitably suffer from drift during long-term usage, while LiDARs provide stable global positioning but yield coarse local positions and orientations, HSC4D mutually complements the two sensors via joint optimization, achieving promising results in long-term capture sessions. We additionally explore the relational cues between humans and their environments to enhance the realism of their interactions. To facilitate a wide range of downstream tasks such as AR, VR, robotics, autonomous driving and more, we present a dataset that includes three large-scale scenes (1k–5k m²) with precisely captured dynamic human subjects.

搜集汇总

背景与挑战

背景概述

HSC4D是一个专注于动态数字世界创建的4D场景数据集，通过车身IMUs和激光雷达的无约束采集方式，提供了包含多样化人类运动和环境互动的数据。该数据集适用于AR、VR、机器人等多个领域，包含三个大场景，旨在支持精准的动态人类行为研究。