基于运动约束的无监督单目视觉里程计  

Unsupervised Monocular Visual Odometry Based on Motion Constraints

作　　者：计志威 刘安东[1] 付明磊[1] 孙虎 张文安[1] 金锦峰 Ryhor Prakapovich Uladzislau Sychou JI Zhiwei;LIU Andong;FU Minglei;SUN Hu;ZHANG Wenan;JIN Jinfeng;Ryhor Prakapovich;Uladzislau Sychou(College of Information Engineering,Zhejiang University of Technology,Hangzhou Zhejiang 310023,China;The United Institute of Informatics Problems of the National Academy of Sciences of Belarus,Minsk 220012,Belarus)

机构地区：[1]浙江工业大学信息工程学院,浙江杭州310023 [2]白俄罗斯国家科学院信息学问题联合研究所,白俄罗斯明斯克220012

出　　处：《传感技术学报》2024年第9期1545-1554,共10页Chinese Journal of Sensors and Actuators

基　　金：国家自然科学基金中白合作交流项目(62111530299)。

摘　　要：针对视觉里程计易受到动态物体及遮挡因素影响导致的位姿估计准确度低、鲁棒性差等问题,提出一种基于运动约束的无监督视觉里程计算法。首先通过最小重投影误差方法处理了大范围场景中的前景遮挡,考虑到现实场景中的运动物体影响,结合光流估计设计了一种运动掩膜处理方法,有效剔除了场景中动态物体像素信息;其次,对于场景中的重复结构和均匀纹理区域,通过数据驱动从轨迹数据中学习车辆的行为模式,建立运动学约束,将车辆运动模型建模为多源时间序列回归模型,避免陷入局部解;最后,结合运动掩膜与运动学模型约束所设计的无监督深度学习框架,对单目相机运动位姿及场景深度进行同步估计,提高了位姿估计精度及模型适应性。在KITTI道路公开数据集和校园低速无人车平台上的实验结果表明,所设计算法的位姿估计精度及深度估计精度优于目前主流无监督单目视觉里程计方法。Targeting at the problems of low accuracy and poor robustness of pose estimation caused by the visual odometry being suscep-tible to dynamic objects and occlusion factors,an unsupervised visual odometry method based on motion constraints is proposed.Firstly,the foreground occlusion in a wide range of scenes is handled by the minimum reprojection error method,and a motion mask processing method is designed in combination with optical flow estimation considering the influence of moving objects in realistic scenes,which ef-fectively eliminates the dynamic object pixel information in the scenes.Secondly,for the repetitive structure and uniform texture regions in the scenes,the kinematic constraints are established by learning the behavior pattern of vehicles from the trajectory data through data-driven approach.Finally,the unsupervised deep learning framework designed by combining the motion mask and the kinematic model constraint is used to estimate the monocular camera motion pose and scene depth simultaneously,which improves the pose estimation ac-curacy and model adaptability.Experimental results on the KITTI road public dataset and the campus low-speed unmanned vehicle plat-form show that the designed algorithm outperforms the current mainstream unsupervised monocular vision odometry methods in terms of positional estimation accuracy and depth estimation accuracy.

关 键 词：视觉里程计 无监督学习 深度估计 位姿估计 运动约束 

分 类 号：TP242[自动化与计算机技术—检测技术与自动化装置]