毕业生知识库

近年来，随着智能机器人的不断发展，其所需完成的任务也愈加复杂。在执行各种复杂任务的过程中，机器人常常需要应对各种复杂场景。这些场景中存在着各种形状和外观各异的动态物体，并伴随着各种无法预测的动态干扰。因此，如何提高机器人对环境的感知能力以及如何准确地建模周围的环境，成为机器人面临的主要问题。本文针对动态场景中的运动物体感知和场景重建展开研究，主要内容如下：

一、介绍了动态场景下运动物体感知与场景重建的研究背景与研究意义。从动态物体感知、不确定动态遮挡下的三维建模和立体匹配三个方面对研究现状进行了综述，并对论文的内容和结构进行了介绍。

二、针对现有运动分割方法对动态物体敏感性较低的问题，提出了一种基于分层采样和位姿流形上概率场模型的动态刚体检测与运动估计方法。首先，在位姿流形上建立概率场，采样过程中区域概率的增长反映了采样的重要性，通过对重要位置的多次采样，提高了动态物体感知的灵敏度。其次，针对采样随机性较大的问题，设计了基于高概率区域的分层采样策略，对每个高概率区域对应的特征点集进行重采样，从而筛选出更多有效的点，减小了采样随机性。最后，在KITTI、KT3DMoSeg、Hopkins 155和MTPV62等数据集上的测试表明了所提方法的有效性。

三、针对工业场景中工件的三维建模无法很好地应对动态遮挡的问题，提出了一种基于双目相机的不确定动态遮挡下的三维建模方法。首先，针对动态遮挡问题，提出了一种基于位姿假设聚类的动态物体分割方法。该方法不依赖物体的先验信息，通过聚类随机采样得到的模型假设获取运动分割，避免了因遮挡物类型未知及运动复杂性等带来的问题。其次，针对分割后待建模物体不完整点云的拼接问题，提出了一种基于重叠视野区域的动态物体局部约束和全局闭环约束的优化方法，保证了三维建模的精度。最后，设计并搭建了实验平台，在该平台上的三维建模实验验证了所提方法的有效性。

四、针对现有SGM多路径解选择方法未考虑全局信息的问题，提出了一种基于对极线结构多模型拟合的立体匹配方法。首先，考虑到极线上像素视差分布的空间几何特性，提出了一种基于多模型拟合的SGM多路径解选择方法。该方法整合了整个极线的结构信息，实现了具有相同特性视差点的自下而上的分层聚类与SGM多路径解的选择。其次，在多模型拟合过程中，针对点集合并与模型选择对点集分布不敏感的问题，提出了一种基于几何鲁棒信息准则与模型假设聚类的多模型拟合方法，通过分析模型参数的空间分布，实现了模型的正确选择，提升了鲁棒性。最后，在KITTI 2012、KITTI 2015、Middlebury数据集上的实验表明了所提方法的有效性。

五、针对运动物体感知与场景重建问题，建立了面向动态场景的机器人定位与场景重建系统。首先，针对实际场景中动态物体位置对感知效率的影响问题，提出了一种增量式感知策略。该策略在所提动态物体挖掘方法的基础上，通过挖掘-剔除-挖掘迭代的方式来分割动态物体。其次，针对静态场景重建，设计了先剔除后拼接的策略，将单帧点云中动态物体点云剔除后再进行拼接，从而避免了现有方法在深度滤波时将动态物体遮挡区域滤除导致的重建不完整问题。最后，室内动态场景实验验证了所设计系统的可行性和有效性。

六、对本文工作进行了总结，并指出了需要进一步开展的研究工作。

In recent years, the demand for intelligent robots has grown, resulting in an expectation for them to complete increasingly complex tasks. To successfully undertake such tasks, robots often encounter various complex scenarios that comprise dynamic objects with unpredictable disturbances in terms of shape and appearance. Therefore, there has been a pressing need to enhance the robot's perception of the environment so that it can accurately model its surroundings. This thesis delves into research focused on the perception of moving objects and scene reconstruction in dynamic environments. The main contents are as follows.

Firstly, the research background and the significance of this thesis are given. It reviews the development of dynamic object perception, 3D modeling under uncertain dynamic occlusions, and stereo matching. The content and structure of this thesis are introduced.

Secondly, aiming at the low sensitivity of existing motion segmentation methods to dynamic objects, a dynamic rigid body detection and motion estimation method based on hierarchical sampling and the probabilistic field model on the pose manifold is proposed. The probabilistic field is established on the pose manifold, and the growth of region probabilities in the sampling process reflects the importance of sampling. Through multiple sampling at important positions, the sensitivity of dynamic object perception is improved. For the problem of high randomness in sampling, a hierarchical sampling strategy based on high-probability regions is designed to resample the feature point sets corresponding to each high-probability region, so as to screen out more effective points. Notably, the proposed method requires no prior information about the scene and motion constraints. Experiments on the KITTI, KT3DMoSeg, Hopkins 155, and MTPV62 datasets demonstrate the effectiveness of the proposed method.

Thirdly, to overcome the challenge in industrial scenes where the 3D modeling of workpieces fails to handle dynamic occlusion properly, a 3D modeling method based on binocular cameras under uncertain dynamic occlusion is proposed. Initially, a dynamic object segmentation method based on pose hypothesis clustering is proposed. This method does not rely on the prior information of the object, and performs motion segmentation via clustering the model hypotheses generated through random sampling, which resolves issues caused by unknown dynamic objects and complex motion patterns. Furthermore, an optimization method based on local constraints of dynamic objects in co-visibility regions and global loop-closing constraints is proposed, aimed at registering the incomplete point clouds of objects that are to be modeled. This method guarantees the accuracy of 3D modeling. Moreover, an experimental platform is designed and built. The 3D modeling experiments conducted on this platform demonstrate the effectiveness of the proposed method.  

Fourthly, the existing SGM multi-path solution selection methods do not utilize the global information. To address this issue, a stereo matching method based on multi-model fitting of the epipolar line structure is proposed. According to the spatial geometric characteristics of the disparity distribution of the pixels on the epipolar line, the multi-model fitting method is adopted to implement the SGM multi-path solution selection. It integrates the structural information of the entire epipolar line. This method also achieves the bottom-up hierarchical clustering of disparity points that have same characteristics and selects SGM multi-path solutions. Furthermore, a multi-model fitting method based on GRIC and model hypothesis clustering is proposed to deal with the problem that the combination of point sets and model selection is not sensitive to the distribution of point sets. It analyzes the spatial distribution of the model parameters to achieve accurate model selection, and thus improves the robustness. Experiments on KITTI 2012, KITTI 2015, and Middlebury datasets demonstrate the effectiveness of the proposed method.

Fifthly, a robot localization and scene reconstruction system for dynamic environments is established. An incremental perception strategy is first proposed to enhance the perception efficiency of dynamic objects. It segments the dynamic objects through mining-culling-mining iterations based on the proposed dynamic object mining method. To reconstruct the static scene, a culling-splicing strategy is designed. The dynamic object point cloud is culled from the single-frame point cloud before splicing. This design helps to avoid the incomplete map problem caused by filtering out the occluded regions of dynamic objects during depth filtering in the existing methods. The feasibility and effectiveness of the designed system are verified through experiments conducted in indoor dynamic environments.

Finally, the conclusions are provided and future work in this area is addressed.