System outliers constitute a different problem than system noise, and frequently even a more crucial one. In our example of the dead animal above, at times the system might generate a single outlier: one data point arbitrarily far away from a sequence of identical data points. Such outliers might have, of course, a devastating effect on the total distance traveled, the speed, and many other possible measures. They have, moreover, a devastating effect on most smoothing methods, which are used to deal with the previous problem of the regular system noise. Intuitively speaking, the difference between noise and outliers is that outliers are something you would have immediately corrected with no scruples whatsoever, if you just happened to notice them. When the data are automatically tracked and analyzed, however, you might not be able to look for them and notice many of them.
This problem is solved in SEE Path Smoother by using robust methods. Robust LOWESS is robust in the sense that it automatically downweights outliers. This is achieved by using the result of one smoothing iteration, as described above, only as the first estimation needed to discover outliers. In a second iteration, this smoothing process is computed again, but this time data points are also weighted according to the absolute difference between them and the first estimation - their estimated error. Data points with higher estimated errors get lower weights, and those with estimated error higher than 6 times the median absolute size of deviations (from the smooth) are considered “outliers” and get a zero weight, i.e., they are completely disregarded from computing the new estimation within each time window. Robust LOWESS is thus used in SEE Path Smoother to solve both the noise problem and the outliers problem, for both location and speed.