Automatic Point Cloud Registration for Large Outdoor Scenes Using a Priori Semantic Information

As an important and fundamental step in 3D reconstruction, point cloud registration aims to find rigid transformations that register two sets. The major challenge techniques is finding correct correspondences the scenes may contain many repetitive structures noise. This paper primarily concerned with improving using a priori semantic information search for correspondences. In particular, we present new pipeline large, outdoor takes advantage of segmentation. Our method consisted extracting segments from clouds efficient deep neural network, then detecting key points feature descriptor get initial correspondence set, and, finally, applying Random Sample Consensus (RANSAC) strategy estimate align same labels. Instead all global alignment, our aligned calculated by each segment highest inlier ratio. We evaluated on publicly available Whu-TLS data set. These experiments demonstrate how improves terms precision speed.