The Worldwide Journal of Robotics Analysis, Volume 43, Issue 8, Web page 1250-1278, July 2024.
Visible movement estimation is a well-studied problem in autonomous navigation. Current work has centered on addressing multimotion estimation in extremely dynamic environments. These environments not solely comprise a number of, advanced motions but in addition are likely to exhibit vital occlusion. Estimating third-party motions concurrently with the sensor egomotion is troublesome as a result of an object’s noticed movement consists of each its true movement and the sensor movement. Most earlier works in multimotion estimation simplify this downside by counting on appearance-based object detection or application-specific movement constraints. These approaches are efficient in particular purposes and environments however don’t generalize effectively to the total multimotion estimation downside (MEP). This paper presents Multimotion Visible Odometry (MVO), a multimotion estimation pipeline that estimates the total SE(3) trajectory of each movement within the scene, together with the sensor egomotion, with out counting on appearance-based data. MVO extends the normal visible odometry (VO) pipeline with multimotion segmentation and monitoring strategies. It makes use of bodily based movement priors to extrapolate motions by short-term occlusions and establish the reappearance of motions by movement closure. Evaluations on real-world information from the Oxford Multimotion Dataset (OMD) and the KITTI Imaginative and prescient Benchmark Suite display that MVO achieves good estimation accuracy in comparison with related approaches and is relevant to a wide range of multimotion estimation challenges.
发布者：Kevin M. Judd，转转请注明出处：https://robotalks.cn/multimotion-visual-odometry/
