The main objective of this work is to search for a motion estimation technique for vehicles and pedestrians having the following properties:

• It should produce estimations with a long Time Horizon.
• It should be as general as possible and work with many diJerent kinds of objects.
• It should be fast enough to give estimations in real time.

We propose a motion estimation technique based on pairwise clustering which verifies the required properties.

We have implemented and tested our approach, comparing it with a technique that we consider to represent the state of the art in clustering techniques. In order to perform the comparison, we propose a benchmark that can be used to test other motion estimation techniques.