基于显著场景Bayesian Surprise的移动机器人自然路标检测

Natural Landmark Detection of Mobile Robots Based on Bayesian Surprise of Salient Scenes

移动机器人在未知非结构化环境下的自然路标检测是层次化环境建模的基础.文中提出一种基于显著场景Bayesian Surprise的自然路标检测方法.通过计算场景的视觉注意图,引导SURF特征采样聚集在显著区域内,提出融合空间关系的词袋模型构造场景表观的模式向量,建立基于该特征描述的地点Multivariate Polya模型,并通过度量传感器观测的Surprise来获取显著场景对应的路标.实验验证自然路标检测方法在大规模复杂室内环境中具有较低的漏检率和误检率,结合层次化SLAM方法验证路标检测器对生成拓扑节点的有效性.

Nature landmark detection of mobile robot in unknown and unstructured environment is a basis of hierarchical environmental mapping. A natural landmark detection method is proposed based on Bayesian Surprise of salient scenes. Visual attention map of scene images is computed to guide the SURF feature sampling within the scope of salient regions. The improved spatial bag-of-words model （sBoW） is employed to construct the pattern vectors of scene appearance. Multivariate Polya model based on the spatial bag-of-words paradigm is proposed for representing the place, and the detection of landmarks corresponding to salient scenes is achieved by computing the surprise of sensor measurements. The experimental results validate the low miss alarm rate and false alarm rate of the nature landmark detection method in large-scale and complex environment, as well as the effectiveness of generating topological nodes with the combination of hierarchical SLAM method.