基于特征编码和动态路由优化的视觉定位方法

Visual localization method based on feature coding and dynamic routing optimization

针对视角变化、光照变化、大尺度和动态物体等复杂场景下,移动机器人定位的准确性低、鲁棒性差等问题,提出基于特征编码和动态路由优化的视觉定位方法。首先,引入基于残差网络的特征编码策略,提取图像的几何特征和语义特征,减少图像噪声信息,加快模型的收敛速度;其次,通过熵密度峰值优化网络的动态路由机制,采用向量表示特征之间的空间位置关系,提升图像特征提取和表达能力,优化网络整体性能;最后,融合优化后的特征编码和动态路由网络,将全局特征描述符和特征向量相结合,保留特征间的差异性和关联性,计算图像特征的相似性用于闭环检测。实验结果表明,相比基于VGG、AlexNet、BoVW及GIST的视觉定位方法,所提方法的准确率分别提高了24.54%、23.06%,43.81%和42.69%,实现了复杂场景下移动机器人闭环检测,提高了定位和建图的准确性和鲁棒性。

To deal with the poor precision and robustness of mobile robot localization in complicated environment such as lighting variations,view variations,large scale and dynamic objects,the visual localization method based on feature coding and dynamic routing optimization is proposed.Firstly,feature coding based on residual network is fused,which can extract semantic and geometric features,remove the noise information of the images,and speed up the convergence.Then,the peak entropy density is used to optimize the dynamic routing of network,the spatial relationship of features is described by vectors,the network performance and precision of image feature extraction are upgraded and the whole performance of the network is optimized.Finally,the optimized feature coding is combined with dynamic routing network,and feature vectors and global descriptors are fused.The images similarity is calculated to realize closed-loop detection,so as to preserve the correlation and difference of features.The experimental results show that compared with the visual localization methods based on VGG,AlexNet,BoVW and GIST,the precision of the proposed method is improved by 24.54%,23.06%,43.81%and 42.69%,respectively.The visual localization method can realize the closed-loop detection of mobile robots in complicated environment and increase the robustness and precision of simultaneous localization and mapping.