A variety of setups and paradigms are used in the neurosciences for automatically tracking the location of an animal in an experiment and for extracting features of interest out of it. Many of these features, however, are critically sensitive to the unavoidable noise and artifacts of tracking. Here, we examine the relevant properties of several smoothing methods and suggest a combination of methods for retrieving locations and velocities and recognizing arrests from time series of coordinates of an animal's center of gravity. We accomplish these by using robust nonparametric methods, such as Running Median (RM) and locally weighted regression methods. The smoothed data may, subsequently, be segmented to obtain discrete behavioral units with proven ethological relevance. New parameters such as the length, duration, maximal speed, and acceleration of these units provide a wealth of measures for, e.g., mouse behavioral phenotyping, studies on spatial orientation in vertebrates and invertebrates, and studies on rodent hippocampal function. This methodology may have implications for many tests of spatial behavior.