服务机器人室内环境层次化地图构建方法研究

发布时间：2018-07-26 19:03

【摘要】：服务机器人感知与理解室内环境是人工智能领域的基础与长远目标之一,因为服务机器人只有了解了人所在的环境,才能实现人与机器人自然交互。一般而言,人与机器人的交流大多是人对机器人发送任务,然后机器人执行人的任务。人类采用自然语言提出的任务,是一种隐含了许多复杂信息的语义关系。隐含的信息包括:实物地点信息、实物的描述信息以及实物之间的关系等等,这些信息分别对应着度量地图、特征地图以及语义地图。单一的地图很难满足机器人实现人类的自然语言提出的任务,所以,构建一种综合性室内地图对于服务机器人理论研究和实际应用均具有重要意义。然而不同种类的地图不能简单的合并,层次化地图是合并地图的一种有效方法,这种合并通过地图级联实现,所以构建一种层次化地图是本文研究重点。为实现服务机器人层次化地图构建,本文针对度量地图、特征地图及语义地图的构建技术分别开展研究:度量地图是层次化地图的基础,而位姿估计又是度量地图构建的基础。本文提出一种尺度加权范数(PMWN,Polar Metric-Weighted Norm)实现2D扫描数据的匹配,解决匹配过程中服务机器人旋转与平移带来的影响,从而减少迭代次数以提高效率。通过对PMWN的残差估计计算出服务机器人的旋转,通过PMWN的几何意义计算出服务机器人的平移。最后,采用基于KF(Kalman Filter)融合里程计得到的位姿与PMWN得到的位姿构建度量地图。特征地图是层次化地图的中间层。本文提出基于假设检验的方法划分度量地图层提供的2D扫描数据,并在数据划分的基础上提取角点、线段以及圆弧特征。采用假设检验的方法匹配这些特征,并采用基于权重的特征融合方法融合特征。同时为了提高度量地图构建的效果,使用这些特征实现闭环检测。最后,通过特征提取、匹配与融合构成特征地图。语义地图是层次化地图的顶层。本文提出了一种语义区域用于表示语义地图。语义区域由语义标识构成,语义标识中包含室内环境的局部轮廓信息,通过匹配轮廓信息采集语义标识。采用粒子滤波器更新语义标识权重,然后通过Viterbi算法估计服务机器人当前所在区域。同时采用基于贝叶斯网络的语义区域导航算法引导服务机器人到达指定区域。层次化地图由度量地图、特征地图及语义地图构成。层次化地图通过地图级联的方式封装服务机器人任务,以实现自然语言描述的服务机器人任务。如果室内环境中出现小范围的改动,那么在层次化地图上可以灵活地变更各层次之间的关系以适应不变的任务。本文首先分别对度量地图、特征地图以及语义地图进行实验验证,然后通过综合实验对层次化地图进行验证。每种地图的实验方法不同,度量地图基于PMWN姿估计实现,通过与PSM及Mb-ICP相比较,验证本文提出算法的准确性、效率及闭环精度;特征地图由特征构成,特征的稳定性与准确性依靠复杂环境特征提取实验及大规模环境的特征提取实验加以验证;语义地图采用语义区域实现,通过长期性实验与机器人绑架实验验证机器人在语义区域上执行长期任务能力。最后层次化地图通过环境变更但任务不变的实验验证层次化地图的灵活性;通过服务机器人在大范围环境中长期执行重复任务的实验验证服务机器人在层次化地图中执行任务的稳定性。
[Abstract]:A service robot perceiving and understanding the indoor environment is one of the basic and long-term goals in the field of artificial intelligence. Because the service robot can only realize the natural interaction between the man and the robot only by understanding the environment of the human being, in general, the communication between human and robot is mostly the task of the human being to the machine man, and then the task of the robot. The task proposed by human language is a semantic relation that implies a lot of complex information. The hidden information includes the physical location information, the description information of the object and the relationship between the objects, etc. the information corresponds to the metric map, the feature map and the semantic map. The single map is difficult to satisfy the robot. Therefore, it is of great significance to build a comprehensive indoor map for the theoretical research and practical application of service robots. However, different types of maps can not be simply merged. The hierarchical map is an effective method for merging maps. This combination is realized by a cascade of maps, so it is constructed. Building a hierarchical map is the key point of this paper. In order to realize hierarchical map building of service robots, this paper studies the construction techniques of metric maps, feature maps and semantic maps. The measurement map is the basis of hierarchical map, while position and posture estimation is the basis of measuring map building. This paper proposes a scale weighting. PMWN (Polar Metric-Weighted Norm) realizes the matching of 2D scanning data and solves the influence of the rotation and translation of the service robot in the matching process, thus reducing the number of iterations to improve the efficiency. The rotation of the service robot is calculated by the residual estimation of the PMWN, and the translation of the service robot is calculated through the geometric meaning of the PMWN. Finally, using the position and posture obtained by the KF (Kalman Filter) fusion odometer and the position and posture obtained by the PMWN, the feature map is the middle layer of the hierarchical map. This paper proposes a method to divide the 2D scan data provided by the metric map layer based on the hypothesis testing, and extract the corner, line segment and arc on the basis of the data partition. We use the method of hypothesis testing to match these features, and use the weight based feature fusion method to fuse the features. In order to improve the effect of the measurement map construction, these features are used to realize closed loop detection. Finally, feature maps are formed by feature extraction, matching and fusion. The semantic map is the top layer of the hierarchical map. A semantic region is used to represent semantic maps. The semantic region is composed of semantic identities. The semantic identification contains the local contour information of the indoor environment. The semantic identification is collected by matching the contour information. The particle filter is used to update the semantic identification weight, and then the current region of the service robot is estimated by the Viterbi algorithm. The semantic regional navigation algorithm based on Bayesian network is used to guide the service robot to the specified area. The hierarchical map is composed of the metric map, the feature map and the semantic map. The hierarchical map is used to encapsulate the service robot task through the cascade of maps to realize the service robot task of the natural language description. If the indoor environment is in the indoor environment There is a small change in it, so the relationship between different levels can be changed flexibly to adapt to the unchanging task on the hierarchical map. First, this paper validates the measurement map, the feature map and the semantic map respectively, and then verifies the hierarchical map through the comprehensive experiment. The experimental methods of each map are different, The metric map is based on PMWN attitude estimation. By comparing with PSM and Mb-ICP, the accuracy, efficiency and close loop accuracy of the algorithm are verified. The feature maps are composed of features. The stability and accuracy of the features depend on the experiment of complex environment feature extraction and the feature extraction experiments of large-scale environment. Semantic map uses semantics. Regional implementation, through long-term experiment and robot kidnapping experiment, verifies the ability of the robot to perform long term task in the semantic area. Finally, the hierarchical map verifies the flexibility of the hierarchical map through the environment change but the task invariable experiment. Service robots perform the stability of tasks in hierarchical maps.
【学位授予单位】：哈尔滨工业大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP242

【相似文献】