A Machine Learning Approach to Track Association

Matthew Schofield

Track association is the process of from a data set of individual spatial temporal data points belonging to many unidentified objects, identifying points belonging to the same object and creating tracks such that all object trajectories represented in the data set are identified. An automated system relying only upon spatial temporal data in order to perform track association is beneficial, as objects one seeks to track would not need additional equipment to report their identity. We employed maritime vessel radar data from a harbor to demonstrate our method. Our method employs multiple iterations, each having a model or heuristic and a target association case, to build upon previous iterations' sub track associations. Scoring track association also involves interesting metrics; completeness, continuity and number of associated tracks. Using these metrics our proposed solution performs very well, even under additionally challenging conditions such as time gaps in the data.