针对同时定位与地图构建(SLAM)的闭环检测方法在多歧义复杂场景下易发生感知偏差的问题, 基于闭环概率模型提出了一种结合局部SURF特征和全局ORB特征的闭环检测方法。首先, 分别采用鲁棒SURF特征和全局ORB特征对图像进行局部和全局的场景描述; 其次, 构建多特征场景描述的离散贝叶斯闭环概率模型, 对多特征空间分别构建观测似然概率, 其中,局部特征空间基于词袋模型的方法计算观测似然概率, 全局特征空间基于KNN最近邻的方法计算观测似然概率; 最后, 考虑图像的时间一致性, 基于极线约束设计多步闭环候选帧提取方法, 进一步减少感知偏差问题。实验结果表明, 在多歧义场景下所提方法可以消除绝大部分的误正匹配情况, 对比FAB-MAP2.0和BoW方法具有更好的闭环检测效果, 可以达到更高的闭环准确率。

The closed-loop detection method for SLAM (Simultaneous Localization and Mapping) is prone to perceptual deviation in complex scenes with multiple ambiguities.Based on the closed-loop probability model,a closed-loop detection method that combines local SURF features with global ORB features is proposed.Firstly,robust SURF feature and global ORB feature are used to describe the image locally and globally.Secondly,the discrete Bayesian closed-loop probability model for multi-feature scene description is constructed,and the observation likelihood probability is constructed for multi-feature spaces,in which the local feature space calculates the observation likelihood probability based on the bag-of-words model,and the global feature space calculates the observation likelihood probability based on KNN nearest neighbor method.Finally,considering the temporal consistency of images,a multi-step,closed-loop candidate frame extraction method is designed based on epipolar constraints to further reduce the perception deviation.The experimental results show that the algorithm can eliminate most of the false-positive matching cases in multi-ambiguity scenes,and has better closed-loop detection effect and higher closed-loop accuracy compared with FAB-MAP2.0 and BoW methods.