Channel Prediction and Target Tracking for Multi-Agent Systems

Doctoral thesis, 2018

Mobile moving agents as part of a multi-agent system (MAS) utilize the wireless communication channel to disseminate information and to coordinate between each other. This channel is error-prone and the transmission quality depends on the environment as well as on the configuration of the transmitter and the receiver. For resource allocation and task planning of the agents, it is important to have accurate, yet computationally efficient, methods for learning and predicting the wireless channel. Furthermore, agents utilize on-board sensors to determine both their own state and the states of surrounding objects. To track the states over time, the objects’ dynamical models are combined with the sensors’ measurement models using a Bayesian filter. Through fusion of posterior information output by the agents’ filters, the awareness of the agents is increased. This thesis studies the uncertainties involved in the communication and the positioning of MASs and proposes methods to properly handle them.

A framework to learn and predict the wireless channel is proposed, based on a Gaussian process model. It incorporates deterministic path loss and stochastic large scale fading, allowing the estimation of model parameters from measurements and an accurate prediction of the channel quality. Furthermore, the proposed framework considers the present location uncertainty of the transmitting and the receiving agent in both the learning and the prediction procedures. Simulations demonstrate the improved channel learning and prediction performance and show that by taking location uncertainty into account a better communication performance is achieved.

The agents’ location uncertainties need to be considered when surrounding objects (targets) are estimated in the global frame of reference. Sensor impairments, such as an imperfect detector or unknown target identity, are incorporated in the Bayesian filtering framework. A Bayesian multitarget tracking filter to jointly estimate the agents’ and the targets’ states is proposed. It is a variant of the Poisson multi-Bernoulli filter and its performance is demonstrated in simulations and experiments. Results for MASs show that the agents’ state uncertainties are reduced by joint agent-target state tracking
compared to tracking only the agents’ states, especially with high-resolution sensors.

While target tracking allows for a reduction of the agents’ state uncertainties, highresolution sensors require special care due to multiple detections per target. In this case, the tracking filter needs to explicitly model the dimensions of the target, leading to extended target tracking (ETT). An ETT filter is combined with a Gaussian process shape model, which results in accurate target state and shape estimates. Furthermore, a method to fuse posterior information from multiple ETT filters is proposed, by means of minimizing the Kullback-Leibler average. Simulation results show that the adopted ETT filter accurately tracks the targets’ kinematic states and shapes, and posterior fusion provides a holistic view of the targets provided by multiple ETT filters.