Stochastic Geometric Information Fusion for Cooperative Driving

The integration between sensors and microprocessors in vehicles has become increasingly tighter with technological advancements. However, each sensor's field of view (FoV) still poses a challenge. Multiple vehicles running local multi-object tracking (MOT) filters can exchange information and fuse it to overcome this. While Bayesian approaches have been the traditional method, newer methods such as random finite set (RFS) approaches, including the probability hypothesis density (PHD) filter and labelled multi-Bernoulli (LMB), demonstrate increased tracking accuracy in complex and dynamic scenarios with multiple targets. This thesis comprehensively explores using RFS-based multi-object tracking filters within an intelligent transport systems (ITS) application. It presents solutions to the significant issues preventing such a system from seeing future adoption. Due to its fundamental advantages, this project advocates using the LMB filter for all ITS scenarios. As centralized methods are optimal but not scalable for large vehicular networks, the usage of distributed networks is explored to create a scalable solution to cooperative information fusion within the ITS framework using multiple LMB filter nodes. The solution to a mathematically correct fusion method is investigated by modifying the generalized covariance intersection (GCI) rule to enable cooperative fusion with limited FoVs. Finally, we present a method to incorporate object class into the LMB filter, allowing superior tracking performance and extending it for use in fusion schemes.