Unsupervised Monocular Visual Odometry Based on Motion Constraints
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.
国家科技期刊平台
NSTL
万方数据
