Summary of Performance and Applicability

• The admissible events of the algorithms are different. Currently, only IP97 can handle all events. The occlusion handling procedure of IP97 can be used in SS90, extending to all events the scope of SS90 as well. The other three algorithms have intrinsic limitations.

• $N$ is a key parameter, with 40 trajectories per $200 \times 200$ size image being a critical value. For higher densities, the performance of most algorithms starts to deteriorate faster.

• The non-iterative HW89, RS91 and IP97 are much faster than the iterative SJ87 and SS90. The difference grows with $N$. The iterative algorithms, especially SJ87, become extremely slow at $N> 40$.

• For $N< 40$, SJ87 offers good trajectory performance at the expense of link performance and computational feasibility.

• HW89 may loose trajectories and give multiple choices. The algorithm is better suited for slow motion and low densities. Otherwise, its performance is suprisingly poor.

• SS90 is less efficient at high speeds and densities. At less severe conditions, the performances of SS90 and IP97 are similar. The trajectory optimization procedure used by SS90 is quite slow, but it may prove useful for coherent trajectories.

• RS91 performs reasonably well at low-to-moderate speeds. At high speeds and densities, its performance deteriorates fast.

• RS91 needs the initial correspondences which limits its application area. In their paper [9], Rangarajan and Shah argue that prior knowledge of the initial correspondences is required for reliable tracking. We experienced that this might only be true when both speeds and densities are high. Otherwise, self-initialization is possible.

• IP97 is the most robust to high speeds and point densities. The speed limit $v_{max}$is the critical parameter that should be set carefully for good operation. This algorithm is an obvious choice for uncorrelated trajectories, high speeds and high point densities.