Picking the raisins out of the cookie LiDAR based odometry in dynamic urban environments

Abstract

Reliable localization using LiDAR on pedestrian-scale delivery robots in urban environments poses distinctive challenges. Methods proven on larger vehicles with high-end sensors do not necessarily translate to compact platforms with simpler payloads. This thesis focuses on evaluating some state-of-the-art LiDAR odometry approaches on a small delivery robot equipped with a LiDAR and other sensors such as an IMU. By comparing trajectory accuracy, resilience to real-world motion profiles, and multiple sensor configurations, insight is provided into which techniques can serve as a reliable baseline and what additional capabilities are required to achieve practical localization on often more resource-constrained robots. Building on this foundation, we explore targeted enhancements driven by real-world observations. We implement a real-time simple dynamic-object filtering approach that operates in parallel with odometry, based on a range images, where moving objects are subsequently identified and excised to maintain a static map. To further mitigate drift over extended runs, we integrate loop-closure adjustments, detecting and aligning previously visited areas to correct accumulated drift over larger distances. Together, these refinements demonstrate how a robust baseline can be augmented to meet the demands of odometry and mapping of a small delivery robot.