论文部分内容阅读
针对移动机器人SLAM(同时定位与地图创建)中的闭环检测问题,提出了一种基于SURF(加速鲁棒特征)和ORB(oriented FAST and rotated BRIEF)全局特征的快速闭环检测算法.首先利用SURF与ORB分别提取查询图像的全局特征,实现对图像的特征表征.在特征提取过程中,对查询图像进行归一化操作,并将归一化的图像中心直接作为SURF与ORB的特征点位置,从而避免了耗时的特征点定位过程.然后将归一化的图像直接作为特征点的邻域区域,把计算的SURF与ORB局部特征描述符作为图像的全局特征.为了融合SURF与ORB全局特征实现查询图像的位置识别,提出了H-KNN(混合K最近邻)的改进算法:WH-KNN(加权混合K最近邻).最后通过跟踪模型实现闭环检测,其核心思想是利用之前闭环检测的结果预测查询图像在地图图像中的位置范围.实验中分别使用采集数据集和牛津大学公开数据集对本文算法进行了验证,同时与传统的BOW(词袋)算法进行了对比.本文算法在两种数据集上分别达到了94.3%和94.5%的准确率,并且查询图像位置识别与全局特征提取的平均时间不到3 ms.其准确性及计算速度都超过了BOW算法,可以准确快速地实现实时闭环检测.
Aiming at the problem of closed-loop detection in SLAM (Simultaneous Localization and Map Creation) of mobile robot, a fast closed-loop detection algorithm based on global features of SURF (accelerated robust feature) and ORB (oriented FAST and rotated BRIEF) ORB to extract the global features of the query image respectively to realize the image feature representation.In the feature extraction process, the query image is normalized and the normalized image center is directly used as the feature point position of SURF and ORB Which avoids the time-consuming process of locating feature points.Then the normalized image is directly used as the neighborhood of the feature points, and the calculated SURF and ORB local feature descriptors are taken as the global features of the image.In order to fuse the global features of SURF and ORB The improved algorithm of H-KNN (Mixed K Nearest Neighbor) is proposed: WH-KNN (Weighted Mixing K Nearest Neighbor) .Finally, closed-loop detection is achieved by tracking model. The key idea is to use the result of previous closed-loop detection Predict the location range of the query image in the map image.In the experiment, we use the collected dataset and the open dataset of Oxford University respectively to validate our algorithm, At the same time, compared with the traditional BOW algorithm, the proposed algorithm achieves the accuracy of 94.3% and 94.5% on the two datasets respectively, and the average time of query image location recognition and global feature extraction is less than 3 ms Its accuracy and computational speed exceed the BOW algorithm, which can realize real-time closed-loop detection accurately and quickly.