bad illumination conditions (these figures have been partially enlarged for better visualization). Thus, the object-tracking module is of great importance to ensure the performance of the perception system. The loop-closure and mapping results are shown in Figure 9, which demonstrates the effectiveness of the proposed deep learning-based, large-scale mapping approach. We build a complete 3D point cloud map of the blue area found in Figure 7. Since we do not have a ground truth to evaluate the accuracy of the map, we compare it only with the CAD drawing to qualitatively demonstrate the mapping performance. We also show six typical places selected by clustering the place descriptors (a) (b) Figure 8. Perception results in the Hong Kong Airport: (a) detection results and (b) tracking results. 1 Typical Place 2 Typical Place Loop Closure Result Comparison With the CAD Drawing 1 3 Typical Place 4 Typical Place 5 Typical Place 6 Typical Place 4 5 2 6 Trajectory and Key Frame Positions 3 3D Point Cloud Map Figure 9. Large-scale loop-closure and mapping results in the Hong Kong Airport. JUNE 2020 * IEEE ROBOTICS & AUTOMATION MAGAZINE * 147