Tracking Agents and Predicting Future Agent Motions via Distributed Multi-Clustered Particle Filtering

This thesis investigates two important problems for intelligent robotic interaction with other agents: (1) object tracking from multiple - and potentially heterogeneous - distributed sensors and (2) predicting future agent motions for interactive robotic navigation. These problems are motivated by the deficiencies of existing mobile robots to navigate amongst humans (or other agents) in an intelligent manner similar to how humans are able to co-navigate: by recognising other agents in the environment, inferring their intentions and planning complementary movement trajectories that lead to efficient joint optimisation for all agents. Many existing mobile robots do not reason about the goal-directed movements of others in the environment, leading to substantial sub-optimality in reaching target locations. In order to address the first problem, we develop PTracking, an algorithm for tracking multiple objects from multiple sensors in a distributed manner using Bayesian filtering (and particle filtering specifically to approximate the generally intractable inference task). The main novelty of the proposed approach is the combination of clustering and mixture models to enable more computationally efficient asynchronous inference. We demonstrate the algorithm’s versatility in a number of realistic applications: robotic soccer, multiple object tracking with mobile sensors, multi-robot surveillance, networked camera tracking of people and maritime surveillance. The second problem has been tackled by employing an Inverse Reinforcement Learning (IRL) approach in combination with PTracking to estimate the reward functions that motivate observed behaviour sequences. A key innovation is that unlike previous IRL methods, which typically assume a fixed state-space representation, the state-space representation is dynamically adapted in the proposed method, so that more modelling emphasis is placed on portions of the space that are frequently visited and less emphasis can be placed on rarely visited portions. This allows significant computational savings versus employing a uniformly detailed state-space representation. We show the benefits of the method for activity forecasting applications, intention prediction and for constructing interactive costmaps to guide robot navigation