Persistent Mapping in Quasi-static Dynamic Environments using Unmanned Aerial Vehicles

Abstract

Mapping and tracking Arctic sea ice is critical to protecting the people and animals that inhabit and transit the Canadian Arctic. Unmanned aerial vehicles (UAVs) are well suited to these activities given the vast area of interest, the remote, potentially dangerous environment, and the need to persistently capture the most current state of the map. The scarcity of reliable global position references at high latitudes means that a UAV must localize itself within a map it is in the process of building -- a problem called Simultaneous Localization and Mapping (SLAM). While traditional SLAM assumes all landmarks are static, UAVs in marine Arctic environments must perform persistent mapping while coping with the motion of ice masses in the dynamic environment. This research proposes piecewise-deterministic quasi-static pose graph SLAM (PDQS-SLAM) to exploit landmark motion information during graph construction. This will contribute to persistent mapping in a dynamic environment.

The primary contribution of this thesis is the relaxation of the static assumption inherent to traditional SLAM algorithms. This is achieved by assigning a kinematic motion model to each landmark, proposing a new loop closure factor structure, and estimating the landmark motion alongside the UAV trajectory in the pose graph. Applicability to a general dynamic environment is achieved by augmenting loop closure edges with a state, governed by a finite state machine (FSM), that captures the edge's behaviour over time. This FSM captures edge behaviours during constant-velocity epochs, and across disrupting events detected using score-based structure learning.

The resulting PDQS-SLAM is validated in both simulations and laboratory experiments. The localization and mapping performance of PDQS-SLAM in a dynamic environment including two mobile landmarks that experience events, is shown to have similar performance when compared to the baseline static SLAM in a static environment. This demonstrates that the static assumption can be relaxed by modelling landmark motion and responding appropriately to inevitable disruptions to that motion. This is a step towards persistent mapping by UAVs in a complex, quasi-static environment inspired by the Canadian marine Arctic.