基于CNN多层特征加权融合的闭环检测算法

Loop closure detection algorithm based on multi-layer feature weighted fusion of convolutional neural network

针对深度学习用于闭环检测中存在空间细节特征丢失的问题,提出基于卷积神经网络(CNN)多层特征加权融合(CNN-F)的闭环检测算法.首先,采用预训练的CNN模型作为特征提取器,从网络中提取输入图像的浅层几何特征和深层语义特征;然后,调整特征图数据进行加权融合,将融合的特征图进行主成分分析(PCA)降维处理后,计算相似性得分用于闭环检测;最后,将算法在数据集New College和City Centre上进行测试.实验结果表明:CNN-F可以有效改善图像特征描述效果,相比于直接使用CNN的闭环检测算法,该算法有更好的准确性和鲁棒性.

To address the problem of missing spatially detailed features in deep learning for loop closure detection,a loop closure detection algorithm based on multi-layer feature weighted fusion of convolutional neural network(CNN-F)was proposed.Firstly,the pre-trained convolutional neural network model was used as a feature extractor to extract the shallow geometric features and deep semantic features of the input image from the network.Secondly,the feature maps were adjusted for the weighted fusion.Then the output fusion features were processed by principal component analysis(PCA)algorithm to calculate the similarity scores for loop closure detection.Finally,the algorithm was tested on the datasets New College and City Centre.The experimental results show that CNN-F can effectively improve image feature description ability which compared to the loop closure detection algorithm using CNN solely.The algorithm has better accuracy and robustness.