光学仪器, 2019, 41 (3): 20, 网络出版: 2019-09-02
基于 VGG16模型的快速闭环检测算法
VGG16 model-based fast loop closure detection algorithm
机器视觉 卷积神经网络 闭环检测 粒子滤波 machine vision convolutional neural network loop closure detection particle filter
摘要
深度卷积神经网络在图像特征表示方面优于传统手工特征, 将其用于闭环检测时还存在计算时间随着数据增长不断增加的问题。为了解决这一问题, 提出了一种基于 VGG16模型的快速闭环检测算法。该算法使用在 ImageNet上预训练的 VGG16网络模型提取图像卷积特征, 并通过一种自适应粒子滤波方法得到闭环候选帧, 以固定运算时间。在主流的闭环检测数据集 City Centre和 New College上对此算法进行测试, 实验结果显示, 该算法在两个数据集上可以分别达到 92%准确率下 70%召回率和 96%准确率下 61%召回率, 超过了同类算法, 并有效解决了计算时间增长的问题。
Abstract
Deep convolution neural network is superior to traditional manual features in the image feature representation, but it still has the problem that the calculation time increases with the increase of data in the loop-closure detection. In order to solve this problem, a fast closed-loop detection algorithm based on VGG16 model is proposed. The algorithm uses the VGG16 network model, which is pre-trained on ImageNet, to extract the image convolution features, and a loop-closure candidate frame is obtained by an adaptive particle filtering method to fix the operation time. The algorithm is tested in the mainstream loop-closure detection dataset City Centre and New College, and the experimental results show that the algorithm can achieve 70% recall rate under 92% accuracy and 61% recall rate under 96% accuracy on the two datasets, which exceeds the conventional algorithms, and effectively solves the problem of calculating time growth.
张学典, 顾璋琦, 秦晓飞. 基于 VGG16模型的快速闭环检测算法[J]. 光学仪器, 2019, 41(3): 20. ZHANG Xuedian, GU Zhangqi, QIN Xiaofei. VGG16 model-based fast loop closure detection algorithm[J]. Optical Instruments, 2019, 41(3): 20.