一种基于全局-局部联合二进制特征的快速闭环检测算法

Fast Closed Loop Detection Algorithm Based on Global-local Joint Binary Features

针对小型机器人视觉SLAM任务中对于精度和实时性的双重需求,本文提出了一种基于全局-局部联合二进制特征的快速闭环检测算法.首先,将查询图像统一为固定大小的正方形,利用局部差异二值算子(Local Difference Binary,LDB)分别提取正方形图像的全局特征和局部特征并进行存储,其中局部特征的关键点由增强的FAST算法提供.然后,针对大尺度场景数据库中全局特征暴力搜索时间过长的问题,引入基于汉明距离的增强局部敏感哈希算法(Locality Sensitive Hashing,LSH)以加快全局特征的搜索.其次,使用局部特征对全局特征匹配的结果进行检验,去伪存真,在保证高实时性的同时提高了检测的准确度.最后,为了验证算法的有效性,我们分别在New_College和City_Center数据集上进行了测试,结果表明,在保证检测准确率100%的前提下,召回率分别达到了73%和38%,完成一次闭环检测仅用时15ms.

In order to satisfy the requirements of precision and real-time in the visual SLAM,a fast closed-loop detection algorithm is proposed for small robots based on global-local joint binary features.First,the query image is unified to fixed-size squares,and the local difference binary algorithm is utilized to extract the global binary features and local binary features of the square image which the features are stored.Note that the key points of local features are provided by the enhanced FAST algorithm.Then,to solve the problem of long searching time of global feature violence in large-scale scene datasets,an enhanced locality sensitive hashing(LSH)algorithm is introduced based on Hamming distance to speed up the search for global features.Next,the result of global feature matching is verified by resorting to local feature.By removing false and storing truth,the accuracy of detection is improved while ensuring the high real-time performance.Finally,in order to verify the effectiveness of algorithm,the New_College and City_Center datasets are utilized to test and demonstrate the effectiveness of the developed algorithm.It should be pointed out that,under the premise of guaranteeing the detection accuracy rate of 100%,the recall rates reach 73%and 38%,respectively.And it takes 15ms on CPU.