Vision systems, either biological or machine, are usually overloaded with a vast amount of information received from photoreceptors. However, only a tiny fraction of this information is really relevant to a given task, say motion detection for example. Thus, filtering irrelevant information is the first important processing task in vision systems. The best solution for this problem is to design a task-oriented vision system in which feature extraction is embedded as early as possible in the model. This paper describes the extraction of features which can be used for representing motion trajectories based upon a local visual motion detector known as the template motion detector . Ultimate advantages are a priori knowledge of features which indicates directions of motion and polarities of edges, and a reduction in data.