基于时间一致性优化的动态场景三维重建方法研究

3D reconstruction is the important premise of scene analysis. With the constant promotion of application requirements of intelligence, the technology of 3D reconstruction is developing towards rapidity and dynamic. It is becoming the important research means in the segmental fields of artificial intelligence, e.g., somatosensory interaction, human machine control, augmented reality, etc. Advanced computer vision technology and high performance processor make it possible to reconstruct dynamic scenes based on the temporal consistency analysis of image sequences. This method can overcome the shortcomings of computing depth maps frame-to-frame, e.g., missing data, unreliable data, and occlusion, etc. However, because of the lacks of scene texture and utilization of temporal information, it is hard to realize 3D reconstruction with high quality. To solve the problem, this project takes the indoor dynamic scene as the research object, acquires the depth images in real time using our portable stereo vision system with texture light projection, explores the optimization method of temporal consistency for depth image sequences, and studies the factors affecting the quality of 3D reconstruction, including 3D registration, model updating, dynamic segmentation and reconstruction methods. The research production will provide a valuable theoretical and methodological guidance on the design, develop and application of 3D reconstruction systems in dynamic scenes.

三维重建是场景分析工作的重要前提，随着智能化应用需求的不断提升，三维重建技术正朝着快速、动态的研究方向发展，它日益成为体感交互、人机控制、增强现实等人机交互细分领域研究的重要手段。先进的计算机视觉技术及高性能处理器的出现，使基于图像序列时间一致性分析的动态三维重建成为可能。该方法可以克服传统逐帧计算深度图像时出现的数据缺失、不可靠、以及遮挡等不足。然而，由于场景缺乏纹理信息、现有方法缺少对时序信息的有效使用，从而限制了三维重建的质量。针对这一问题，本项目以室内动态场景为研究对象，利用本室研发的基于纹理光投射的便携式立体视觉系统实时获取深度图像，探索针对深度图像序列的时间一致性优化方法，并对影响三维重建质量的三维配准、模型数据更新、动态分割与重建方法进行研究。项目的研究成果将为面向动态场景的三维重建系统的设计、研发及应用提供有力的理论与方法指导。

本项目旨在以室内动态场景为研究对象，利用本实验室研发的基于纹理光投射的便携式立体视觉系统实时获取深度图像，探索针对深度图像序列的时间一致性优化方法，进而重建出准确的动态场景三维模型。项目组开展了系统而深入的研究工作，建立了三维传感与模式识别相结合的研究体系，取得了一些研究成果。我们成功地解决了有关动态交互场景三维建模中的一些关键问题，实现了准确地对动态物体进行三维建模：（1）开发了动态交互场景三维点云配准方法；（2）开发了交互场景深度数据更新、融合方法； （3）开发了快速、准确的动态前景三维建模方法。