Investigating the Added Value of Combining Regression Results from Different Window Lengths

Predicting the motion of a dynamic object by using time series is highly dependent on the chosen window size. If we use a too short window length the prediction gets noisy and is strongly affected by measurement errors. Too large windows lead to an inert behavior. In this paper we want to analyze, if there is an added value of using regression results coming from different window sizes instead of one, for classifying the performed maneuver. For our investigation, we are using motion data from pedestrians, taken by a laser. Those trajectories are the basis for our regression predictions. From those predictions, we are calculating features, based on one regression result with one window size, or based on two results with different window sizes. In the next step, we analyze which features are more significant for maneuver detection. The evaluation is done with feature selection methods like Principal Component Analysis or Extremely Randomized Trees. Additionally, we are training Neural Networks on features sets, stemming from regression results of one, ore two window sizes. On the outcome of those tests we can estimate the added value of two different regression window length. If features, which are using information from short- and longtime regression, are ranked better than features which are only based on one regression output we can conclude that there is an added value by using different window sizes.