融合卷积神经网络与VLAD的闭环检测方法

LOOP CLOSURE DETECTION ALGORITHM BASED ON CONVOLUTIONAL NEURAL NETWORK AND VLAD

传统的闭环检测方法大多采用人工设计的特征,很容易受到环境的影响。卷积神经网络通过提取层次化特征,更好地应对了光照变化,但忽略了图像的局部空间特性。针对该问题,提出一种融合VGG16与VGG-NetVLAD的闭环检测算法。该网络保留了VGG16的部分结构,并在最后一层引入了基于局部聚合描述符向量(VLAD)思想的池化层NetVLAD,使提取的特征更适用于闭环检测。实验表明,相较于传统的视觉词袋模型及其他几种深度学习方法,该算法具有更强的泛化性,可以在闭环检测中达到更高的准确率并满足实时性的要求。

The traditional methods for loop closure detection are vulnerable to environmental influence,because they always adopt features which are low-level and designed artificially.Convolutional neural network can better deal with light changes by extracting hierarchical features,but also ignores the local spatial characteristics of images.In view of this,we proposed a loop closure detection algorithm which combined VGG16 with VLAD(VGG-NetVLAD).This network retained part of the structure of VGG16,and added the pooling layer NetVLAD based on the idea of VLAD to the last layer,making the extracted features more suitable for loop closure detection.The experiment results show that our model can achieve better precision compared with bag-of-words and other deep learning methods,and also has stronger generalization and meet the requirement of real-time performance.