LiDARs have attached much attention from both academics and industry due to their active nature in measuring distance and robustness to lighting changes. The development of LiDAR technology has facilitated the wide applications of robots in complex environments, such as field surveillance, exploration, search-and-rescue, and autonomous driving. To cope with drawbacks of a single LiDAR in data sparsity and limited field of view, combining multiple LiDARs to maximize a robot's perceptual awareness of environments and obtain sufficient measurement is a straightforward but promising solution. But novel algorithms to unlock the potential of multiple LiDARs must be investigated.

In this thesis, I formulate the multi-LiDAR perception problem in a unified way. Three essential problems ranging f...[ Read more ]

LiDARs have attached much attention from both academics and industry due to their active nature in measuring distance and robustness to lighting changes. The development of LiDAR technology has facilitated the wide applications of robots in complex environments, such as field surveillance, exploration, search-and-rescue, and autonomous driving. To cope with drawbacks of a single LiDAR in data sparsity and limited field of view, combining multiple LiDARs to maximize a robot's perceptual awareness of environments and obtain sufficient measurement is a straightforward but promising solution. But novel algorithms to unlock the potential of multiple LiDARs must be investigated.

In this thesis, I formulate the multi-LiDAR perception problem in a unified way. Three essential problems ranging from extrinsic calibration, simultaneous localization and mapping (SLAM), and recognition are addressed by proposing a coherent and complete perception solution using multiple LiDARs. Extensive simulated and real-world experiments on various robotic platforms have demonstrated the performance of this proposed solution.

I start by presenting an offline method that enables automatic dual-LiDAR extrinsic calibration using one-shot measurements. I further investigate the automatic calibration and propose a hybrid approach that takes advantages of motion features for initialization as well as appearance cues for refinement. Based on the above research, I propose a complete system that enables the flexible and online calibration as well as SLAM with multiple LiDARs. After that, I investigate the algorithm latency problem and propose a greedy-based method for feature selection to accelerate the SLAM system. Moreover, I turn to study recognition problems and integrate the calibration methodology with learning approaches to propose a multi-LiDAR 3D object detector with the awareness of extrinsic perturbation for autonomous driving. Finally, I summarize this thesis and propose future research opportunities.