面向物体级的视觉SLAM动态三维场景解析与重建

SLAM (Simultaneous Localization and Mapping), which is an effective technique for automatic localization and 3D scene reconstruction, has been extensively used in the intelligent robots. Current SLAM algorithms assume that the objects to be localized and 3D reconstructed are static in the scene and cannot be directly applicable for dynamically moving objects. Moreover, excessive noise are inevitably involved in the 3D reconstruction stage, which makes the algorithms difficult to be optimized. To overcome these issues, we proposed a novel object-level visual SLAM algorithm based on dynamic 3D analysis and reconstruction. Our proposed algorithm first uses high-level semantic features and then followed by a top-down technique for 3D object detection and optimization. More specifically, our algorithm establishes a motion analysis model to decouple the object motion and camera motion so as to effectively and robustly extract object motion features and camera poses. An object-level inter-frame connection graph is then built for 3D scene reconstruction and optimization. Finally, a loop closure method based on deep learning is designed to further optimize the SLAM results. Our project aims to analyze the motion of highly complicated dynamic scenes for object localization and 3D reconstruction. By overcoming the above mentioned issues of visual SLAM for dynamic scenes, our proposed technique is capable of providing critical algorithms and theoretical foundation for intelligent robots to autonomous navigation under complicated and dynamic scenes.

SLAM（即时定位与地图构建）技术能够实现高效自主定位和三维重建，在智能机器人中得到了广泛应用。当前SLAM技术大多假设物体是静止的，并且在三维建模时会有大量的噪声，优化复杂度高。针对以上两个问题，本项目拟开展面向物体级视觉SLAM的动态三维场景解析与重建研究，主要研究内容如下：利用物体的高层语义特征和自顶向下技术实现三维目标的检测与优化；构建物体运动和相机运动解析模型以提取物体的运动特征与相机的姿态；建立物体级别的帧间关联图，结合光束平差法以实现场景的三维重建并进行优化；设计基于深度学习的闭环检测技术以进一步优化SLAM技术。本项目旨在对复杂动态场景中的运动进行解析，以实现动态场景下的三维重建与优化，解决视觉SLAM技术在复杂动态场景下的应用瓶颈，为智能机器人在复杂环境下的自主导航提供关键技术和理论基础。

本项目围绕即时定位与建图（SLAM）技术中无法适用于动态场景、三维重建易受噪声影响、复杂度高等问题，按项目计划着重研究了三维目标检测与优化、复杂动态场景运动解析、三维重建与三维场景解析和基于深度场景特征闭环检测等内容，取得了较好的理论成果，并将部分成果转化为实际应用产品。项目组顺利地完成了项目计划书中所确定的研究内容和相关成果指标，取得的研究成果具体如下：已经录用和发表论文25篇（其中，发表SCI检索论文17篇，发表EI检索论文3篇，发表中文核心论文5篇）；申请专利13项，其中授权发明专利2项，授权实用新型专利1项，在审发明专利10项；培养硕士研究生22名，其中19名硕士生已完成答辩并顺利毕业，3位在读研究生。本项目的研究对于推动智能机器人在复杂环境下的自主导航等应用具有重要意义。