基于动态遍历的分层特征网络视觉定位

Hierarchical Feature Network for Visual Localization Based on Dynamic Traversal

采用分层特征网络估计查询图像的相机位姿,会出现检索失败和检索速度慢的问题。对分层特征网络进行分析,提出采用动态遍历与预聚类的视觉定位方法。依据场景地图进行图像预聚类,利用图像全局描述符获得候选帧集合并动态遍历查询图像,利用图像局部特征描述符进行特征点匹配,通过PnP算法估计查询图像的相机位姿,由此构建基于MobileNetV3的分层特征网络,以准确提取全局描述符与局部特征点。在典型数据集上与AS、CSL、DenseVLAD、NetVLAD等主流视觉定位方法的对比结果表明,该方法能够改善光照与季节变化场景下对候选帧的检索效率,提升位姿估计精度和候选帧检索速度。

When used to estimate the camera pose of the query image,Hierarchical Feature Network(HFNet)is limited by frequent retrieval failures and the low retrieval speed.This paper analyzes HFNet and proposes a visual location method based on dynamic traversal and pre-clustering.According to the scene map,the image is pre-clustered.Then the global image descriptor is used to obtain the candidate frame set and dynamically traverse the query image,while the local feature descriptor is used to match the feature points.In addition,the camera pose of the query image is estimated by using the PnP algorithm.On this basis,an HFNet based on MobilenetV3 is constructed for the extraction of global descriptors and local feature points.Experimental results on typical data sets show that,compared with mainstream visual localization methods such as AS,CSL,DenseVLAD and NetVLAD,the proposed method can improve the retrieval efficiency of candidate frames in the cases of changing illumination conditions and seasons.It can also improve the accuracy of pose estimation and the speed of retrieving candidate frames.