基于视觉词典的单目视觉闭环检测算法

Loop Closure Detection AlgorithmBased on Monocular Vision Using Visual Dictionary

针对移动机器人单目视觉同步定位与地图构建中的闭环检测问题,文中设计一种基于视觉词典的闭环检测算法.算法对采集的每帧图像通过SURF进行特征提取,应用模糊K均值算法对检测的视觉特征向量进行分类,在线构建表征图像的视觉词典.为精确表征局部视觉特征与视觉单词间的相似关联,利用混合高斯模型建立视觉词典中的每一视觉单词的概率模型,实现图像基于视觉词典的概率向量表示,通过向量的内积来计算图像间的相似度.为保证闭环检测的成功率,应用贝叶斯滤波融合历史闭环检测与相似度信息来计算闭环假设的后验概率分布.另外,引入浅层记忆与深度记忆两种内存管理机制来保证算法执行的快速性.实验结果证明该方法的有效性.

Aiming at the problem of loop closure detection in monocular simultaneous localization and mapping for mobile robots, a detection algorithm based on visual dictionary （VD） is presented. Firstly, feature extraction is performed for each required image using SURF methods. Subsequently, a fuzzy K-means algorithm is employed to cluster these visual feature vectors into visual words based on VD which is constructed online. To precisely represent the similarities between each visual word and corresponding local visual features , Gaussian mixture model is proposed to learn the probability model of every visual word in bags of visual words. Consequently, every image can be denoted as a probabilistic vector of VD, and thus the similarities between any two images can be computed based on vector inner product. To guarantee the continuity of the closed-loop detection, a Bayesian filter method is applied to fuse historical closed-loop detection information and the obtained similarities to calculate the posterior probability distribution of closed-loop hypothesis. Furthermore, two memory management mechanisms, shallow memory and deep memory, are introduced to improve the process speed of the proposed algorithm. The experimental results demonstrate the validity of the proposed approach.