基于HOG特征与卷积神经网络的闭环检测算法

Loop Closure Detection Algorithm Based on HOG Feature and Convolutional Neural Networks

在同时定位与建图(Simultaneous localization and mapping,SLAM)的闭环检测中,由于受到光照变化、季节变化、遮挡、动态物体、视点变化等带来的影响,即使是同一位置拍摄的两张图片,其像素也会发生很大的变化,进而图片描述子也会有很大的不同。相较于卷积特征,传统人工设计特征存在图片表征力不足的问题。文中通过利用方向梯度直方图(HOG)特征的几何信息与卷积特征融合,使提取的图片描述子对于图片视点变化更加鲁棒;在此基础上,使用一种灵活的闭环检测算法,提高算法对于图片外观发生变化的场景下鲁棒性。为了验证算法的性能,在3个不同环境特点的公开数据集上,与其他的先进算法进行试验对比,试验结果表明,文中算法在复杂环境下仍然具有较高准确率和较大的闭环曲线下面积。

In the loop closure detection of SLAM(simultaneous localization and mapping),even two pictures taken at the same location are affected by illumination change,seasonal variation,occlusion,dynamic objects and viewpoint change.The pixels will also differ greatly,and then the picture descriptors will be obviously different.Compared with convolution features,traditional artificial design features have the problem of insufficient image representation.In this paper,the geometric information of HOG(histogram of oriented gradient)feature is fused with the convolution feature to make the extracted image descriptor more robust to the image viewpoint change.On this basis,a flexible loop closure detection algorithm is proposed to improve the robustness of the algorithm to the scene where the image appearance changes.In order to verify the performance of the proposed algorithm,experiments are carried out on three public data sets with different environmental characteristics and compared with other advanced algorithms.Experimental results show that the proposed algorithm still has high accuracy and large area under loop closure curve in complex environment.