针对传统闭环检测算法在动态场景下不稳定、易失败的问题, 提出了一种在动态场景下能够准确检测到闭环的算法。首先, 改进基于场景流区分动、静特征点算法, 使得到的动静点划分更为准确;其次, 剔除动态特征点, 进行聚类, 将图像在视觉词典树各个节点的TF-IDF熵作为图像在该视觉单词的得分权重, 从而构造得分向量对场景进行描述;最后, 采用负指数幂函数作为计算两幅图像的相似性得分函数, 计算当前帧与候选关键帧的相似性得分, 经过最后的闭环确认环节, 得到最终的与当前帧发生闭环的关键帧。实际场景的实验表明, 所提算法在动态场景下能够有效检测到闭环。

To solve the problem that the traditional Loop Closure Detection(LCD)algorithms are unstable and easy to fail in dynamic scenes,an algorithm that can accurately detect the loop closure under dynamic scenes is proposed.First,the algorithm for distinguishing the dynamic from the static features based on the scenes flow is improved so that the dynamic and static points are more accurately determined.Then,the dynamic feature points are removed and clustering processing is performed.The TF-IDF entropy of each node of the image in the visual dictionary tree is used as the weight of the image in the visual word,and a score vector is constructed to describe the scenes.Finally,the negative exponent power function is used as the similarity score function to calculate the similarity scores of the two images.The similarity score between the current frame and the candidate key frame is calculated,and after the final loop closure confirmation,the final key frame forming a loop closure with the current frame is obtained.The experiment of the actual lab scenes shows that the proposed algorithm can effectively detect the loop closure in the dynamic scenes.