基于卷积神经网络的回环检测算法

Loop Closure Detection Based on Convolutional Neural Network

论文主要针对于移动机器人在视觉定位和建图过程中的回环检测问题进行研究。回环检测是视觉SLAM中的一个至关重要的部分之一。在机器人移动过程中,机器人通过估算自身的位姿,以及感知周围的环境,实现定位和建图。由于机器人在估计位姿时使用的是帧间位姿估计,因此位姿的估计回随着时间的增加而产生漂移。回环检测则是针对于解决位姿漂移问题。现在比较流行的方法就是使用人工建立的特征,使用视觉词袋的方法,来实现回环检测。论文提出了一种基于深度学习的卷积神经网络的回环检测方法。移动机器人通过传感器获取视觉图像的数据,将其输入到与训练好的卷积神经网络中,使用卷积特征作为图像的描述,然后对提取的特征进行处理,计算图像的相似度得分。最后在使用了本地的数据集和TUM数据集进行验证算法的有效性。

This paper focuses on the problem of loop closure detection in mobile robots during visual positioning and mapping. Loop closure detection is one of the most important parts of visual SLAM. During the movement of the robot,the robot realizes posi. tioning and mapping by estimating its own posture and sensing the surrounding environment. Since the robot uses the inter-frame pose estimation when estimating the pose,the estimation of the pose is drifted over time. Loop closure detection is aimed at solving the pose drift problem. Nowadays,the more popular method is to use the artificially built features and use the visual word bag meth. od to achieve loopback detection. This paper proposes a loopback detection method based on deep learning for convolutional neural networks. The mobile robot acquires the data of the visual image through the sensor,inputs it into the trained convolutional neural network,uses the convolution feature as the description of the image,and then processes the extracted feature to calculate the simi. larity score of the image. Finally,the validity of the verification algorithm is performed using the local dataset and the TUM dataset.