Incremental Feature Forest for Real-Time SLAM on Mobile Devices
Real-time SLAM is a prerequisite for online virtual and augmented reality(VR and AR)applications on mobile devices.Under the observation that the efficient feature matching is crucial for both 3D mappings and camera locations in the feature-based SLAM,we propose a clustering forest-based metric for feature matching.Instead of a predefined cluster number in the k-means-based feature hierarchy,the proposed forest self-learn the underlying feature distribution,where the affinity estimation is based on efficient forest traversals.Considering the spatial consistency,the matching feature pair is assigned a confident score by virtue of contextual leaf assignments to reduce the RANSAC iterations.Furthermore,an incremental forest growth scheme is presented for a robust exploration in new scenes.This framework facilitates fast SLAMs for VR and AR applications on mobile devices.
Yuke Guo Yuru Pei
Luoyang Institute of Science and Technology,Luoyang,China Key Laboratory of Machine Perception(MOE),Department of Machine Intelligence,Peking University,Beiji
国际会议
广州
英文
431-438
2018-11-23（万方平台首次上网日期，不代表论文的发表时间）