基于二进制生成对抗网络的视觉回环检测研究

PDF下载HTML

Title:

Visual loop closure detection based on binary generative adversarial network

作者:

杨慧, 张婷, 金晟, 陈良, 孙荣川, 孙立宁

苏州大学 机电工程学院，江苏 苏州 215021

Author(s):

YANG Hui, ZHANG Ting, JIN Sheng, CHEN Liang, SUN Rongchuan, SUN Lining

School of Mechanical and Electric Engineering, Soochow University, Suzhou 215021, China

关键词:

回环检测; 无监督学习; 二进制描述符; BoVW; 视觉SLAM; 生成对抗; 特征提取; 深度学习

Keywords:

loop closure detection; unsupervised learning; binary descriptor; BoVW; visual SLAM; generative adversarial; feature extraction; deep learning

分类号:

TP181

DOI:

10.11992/tis.202007007

摘要:

针对现有的回环检测模型大多基于有监督学习进行训练，需要大量标注数据的问题，提出一种视觉回环检测新方法，利用生成对抗思想设计一个深度网络，以无监督学习的方式训练该网络并提取高区分度和低维度的二进制特征。将距离传播损失函数和二值化表示熵损失函数引入神经网络，将高维特征空间的海明距离关系传播到低维特征空间并增加低维特征表示的多样性，进而利用BoVW模型将提取的局部特征融合为全局特征用于回环检测。实验结果表明：相比SIFT和ORB等特征提取方法，所述方法在具有强烈视角变化和外观变化的复杂场景下具有更好的性能，可以与AlexNet和AMOSNet等有监督深度网络相媲美。但采用无监督学习，从根本上避免了费时费力的数据标注过程，特别适用于大规模开放场景的回环检测，同时二进制特征描述符极大地节约了存储空间和计算资源。

Abstract:

In view of the problem that the existing loop closure detection models are mostly trained based on supervised learning and require a large amount of labeled data, this paper proposes a new method for visual loop closure detection. The idea of the generative adversarial network is adopted, and thus, a deep neural network is designed and trained through unsupervised learning methods to extract more discriminative binary feature descriptors with low dimensions. The distance propagation loss function and a binarized representation entropy loss function are introduced into the neural network. The first loss function can help spread the Hamming distance relationship of the high-dimensional feature space to the low-dimensional feature space, and the second one increases the diversity of the low-dimensional feature representation. The extracted local features are fused into global features by using the BoVW model for further loop closure detection. Experimental results show that the proposed method has better performance than feature extraction algorithms such as SIFT and ORB in complex scenes that have a strong viewpoint and appearance changes, and its performance is comparable with that of supervised deep networks such as AlexNet and AMOSNet. It is especially suitable for loop closure detection in large-scale open scenes because the time-consuming and tedious process of supervised data annotation is completely avoided with the use of unsupervised learning. Moreover, the binary feature descriptors can greatly save storage space and computing resources.