In our future work we will address more accurate modeling of the environment. Currently, tracks are mapped onto events with a non-probabilistic map of the environment. This results in high sensitivity of the event generation to subtle changes in timing of the tracker. The work, presented in  suggests that such maps can be learned automatically. We are also planning on learning the rule probabilities, observing the environment for extended period of time. This will help more accurate modeling the traffic patterns as well as performance of the tracker.
We are planning on better utilization of object correspondences. In current implementation partial tracks are joined only based on the object's position and velocity, as reflected by the penalty function in equation 5. In the future, change in object appearance will also be considered in computing this penalty.