Visual Slam Fast Loop Detection Algorithm Based on Lightweight CNN
蒋经纬 1吉月辉 1刘俊杰 1高强1
扫码查看
点击上方二维码区域，可以放大扫码查看
作者信息
1. 天津理工大学电气工程与自动化学院,天津 300384
折叠
摘要
传统基于卷积神经网络(Convolutional Neural Network,CNN)的视觉同步定位与地图构建(Simultaneous Localization and Mapping,SLAM)系统回环检测目前准确率和召回率较高,但其存在特征提取时间较长和特征向量维度过高导致计算量较大等问题.针对上述问题,结合轻量级卷积神经网络MobileNetV3 和PCA降维算法,提出了一种基于深度学习的快速回环检测算法.基于MobileNetV3 进行特征提取并构建特征矩阵,运用PCA降维算法完成降维以提升运行速度,使用余弦相似度计算各个特征向量间的相似性,并取最大值与给定阈值比较判断是否构成回环.最后,使用New College和City Centre两个公开的数据集验证算法的性能.实验结果表明,相较于传统的CNN回环检测方法,提出的算法在保证准确率和召回率的同时,运行速度更快,较好的满足了视觉SLAM系统准确性和实时性的要求.
Abstract
The traditional visual synchronous localization and mapping(SLAM)system based on Convolutional Neural Network(CNN)has high accuracy and recall in loop detection.However,it has some problems such as long feature extraction time and high dimensionality of feature vectors,resulting in high computational complexity.To solve these problems,a fast loop detection algorithm based on deep learning is proposed by combining the lightweight convo-lutional neural network mobilenetv3 and the PCA dimension reduction algorithm.Feature extraction is carried out based on mobilenetv3 and the feature matrix is constructed.PCA dimension reduction algorithm is used to reduce the dimension and improve the running speed.Cosine similarity is used to calculate the similarity between each feature vector and take the maximum value.It is compared with the given threshold to determine whether a loop is formed.Fi-nally,two public datasets of New College and City Centre are used to verify the performance of the algorithm.The ex-perimental results show that compared with the traditional CNN loop detection method,the algorithm not only guaran-tees accuracy and recall,but also runs faster,which better meets the requirements of the accuracy and real-time of the visual slam system.
维普
万方数据