基于卷积神经网络的视觉位置识别方法

摘要

Aiming at the problem that the traditional methods of place recognition based on hand-crafted features are easily influenced by appearance changes, such as lighting change, seasonal variation and so on, an algorithm for place recognition based on convolutional neural networks (CNN) was proposed.Based on the structure of CNN, a method was designed to extract image descriptors and to improve the robustness in terms of appearance changes.The Bag-of-Words model was trained offline on an open dataset to speed up the process of image matching.The matching result of place recognition was obtained by calculating the similarity score of two images.Experimental results show that the proposed method is more robust for visual place recognition in the condition of appearance changing.%针对传统基于手工特征的视觉位置识别方法受光照、季节等场景外观变化影响较大的问题, 提出一种基于卷积神经网络的位置识别算法.根据卷积神经网络的网络结构设计一种图像描述符提取方法, 提高所提取图像特征的鲁棒性;使用公开数据集离线训练视觉词袋模型, 利用该模型加快图像匹配过程的搜索速度;通过计算两幅图像的相似度分数, 得到位置识别的匹配结果.实验结果表明, 该算法在场景外观变化的情况下位置识别的鲁棒性优于传统方法.