Deep Localization of Static Scans in Mobile Mapping Point Clouds

Abstract

Mobile laser scanning (MLS) systems are often used to efficiently acquire reference data covering a large-scale scene. The terrestrial laser scanner (TLS) can easily collect high point density data of local scene. Localization of static TLS scans in mobile mapping point clouds can afford detailed geographic information for many specific tasks especially in autonomous driving and robotics. However, large-scale MLS reference data often have a huge amount of data and many similar scene data; significant differences may exist between MLS and TLS data. To overcome these challenges, this paper presents a novel deep neural network-based localization method in urban environment, divided by place recognition and pose refinement. Firstly, simple, reliable primitives, cylinder-like features were extracted to describe the global features of a local urban scene. Then, a probabilistic framework is applied to estimate a similarity between TLS and MLS data, under a stable decision-making strategy. Based on the results of a place recognition, we design a patch-based convolution neural network (CNN) (point-based CNN is used as kernel) for pose refinement. The input data unit is the batch consisting of several patches. One patch goes through three main blocks: feature extraction block (FEB), the patch correspondence search block and the pose estimation block. Finally, a global refinement was proposed to tune the predicted transformation parameters to realize localization. The research aim is to find the most similar scene of MLS reference data compared with the local TLS scan, and accurately estimate the transformation matrix between them. To evaluate the performance, comprehensive experiments were carried out. The experiments demonstrate that the proposed method has good performance in terms of efficiency, i.e., the runtime of processing a million points is 5 s, robustness, i.e., the success rate of place recognition is 100% in the experiments, accuracy, i.e., the mean rotation and translation error is (0.24 deg, 0.88 m) and (0.03 deg, 0.06 m) on TU Delft campus and Shanghai urban datasets, respectively, and outperformed some commonly used methods (e.g., iterative closest point (ICP), coherent point drift (CPD), random sample consensus (RANSAC)-based method).