深度构建地图：基于多点云的无监督地图估测.pdfVIP

DeepMapping: Unsupervised Map Estimation From Multiple Point Clouds Li Ding Chen Feng l.ding@ cfeng@ University of Rochester NYU Tandon School of Engineering Mitsubishi Electric Research Laboratories (MERL) 8 1 Abstract 0 2 We propose DeepMapping, a novel registration frame- work using deep neural networks (DNNs) as auxiliary v o functions to align multiple point clouds from scratch to a globally consistent frame. We use DNNs to model the N highly non-convex mapping process that traditionally in- (a) Ground truth (b) DeepMapping 8 volves hand-crafted data association, sensor pose initial- 2 ization, and global refinement. Our key novelty is that prop- ] erly defining unsupervised losses to “train” these DNNs V through back-propagation is equivalent to solving the un- C derlying registration problem, yet enables fewer dependen- . cies on good initialization as required by ICP. Our frame- s c work contains two DNNs: a localization network that es- (c) Direct optimization (d) Multiway registration [11] [ timates the poses for input point clouds, and a map net- Figure 1. Global registration results from the AVD [4]. The base- 1 work that models the scene structure by estimating the oc- line method in (c) is explained in Section 4.1 and 4.2. Each color v cupancy status of global coordinates. This allows us to con- represents one point cloud. Best viewed in color. 7 vert the registration problem to a binary occupancy clas- 9 sification, which can be solved efficiently using gradient- clouds into a complete