From Geometry-Aware Planning to Cooperative Infrastructure Sensing: Toward Robust Autonomous Mobility

Abstract

Safe autonomous mobility requires perception and motion planning systems that can operate reliably under limited sensor coverage, occlusion, adverse environmental conditions, and communication delay. This thesis investigates this problem in both onboard and infrastructure-supported autonomous mobility. It first develops geometry-aware perception and motion planning methods for autonomous vehicles using onboard sensor data, and then extends the framework to cooperative infrastructure sensing for outdoor vehicle-infrastructure cooperation and indoor robotic mobility.

The first part of the thesis focuses on onboard perception and planning. A point-cloud-based motion planning framework is developed by constructing a potential field directly from LiDAR observations and integrating it into a Model Predictive Control formulation. Instead of relying only on object-level detections or predefined object categories, the method uses geometric drivable-area boundaries extracted from point clouds to support real-time trajectory generation around irregular and previously unseen obstacles. This framework is further elaborated by a multi-modal drivable-space perception method that fuses LiDAR, camera, and HD map information. The perception pipeline combines adaptive ground removal, adaptive clustering, and robust LiDAR-camera association to extract safe drivable areas. The perception module has been validated in real autonomous-shuttle tests, confirming its reliability and efficiency, and simulations demonstrate robust and efficient motion planning under challenging conditions.

The second part of the thesis addresses the limitations of onboard sensing by developing cooperative infrastructure perception systems. For outdoor autonomy, the thesis presents CoInfra, a large-scale cooperative infrastructure perception system and dataset for vehicle-infrastructure cooperation. The system deploys multiple synchronized infrastructure sensor nodes equipped with cameras, LiDAR, edge computing, and 5G communication. A delay-aware fusion strategy is introduced to aggregate multi-node observations into a shared global representation while accounting for communication latency and late-arriving data. The resulting dataset provides synchronized multi-modal infrastructure data with globally consistent 3D annotations and diverse weather coverage, supporting evaluation of cooperative perception under realistic urban conditions.

The third part extends cooperative infrastructure sensing to indoor human-centric mobility. A delay-aware indoor perception framework is developed using ceiling-mounted infrastructure sensor nodes, local multi-modal perception, central fusion, and latency-compensated tracking. The system supports real-time estimation of robot, pedestrian, and obstacle states in crowded indoor spaces. An infrastructure-supported 3D human pose estimation method is further developed to improve human position and orientation estimation under occlusion. The fused indoor perception output is integrated with the proposed MPC-based planner and evaluated in closed-loop experiments with a mobile robot platform.

Together, these contributions show how explicit geometric representations, multi-modal sensing, and delay-aware infrastructure perception can improve the robustness and deployability of autonomous mobility systems. By progressing from onboard perception and planning to outdoor and indoor cooperative infrastructure sensing, this thesis provides algorithms, system implementations, datasets, and real-world evaluations for reliable autonomous mobility in complex environments.