Experiments
Data Pre-processing
As with any machine learning device, the input data must contain recurring stable patterns to be learned. One way to track trajectories is to expect patterns based on the environment (streets, walkaways, buildings, trees and objects on the paths), hence image coordinates fix those references for all pedestrians (e.g., at a given set of coordinates there are unavoidable obstacles). However, this choice does not port over to a different dataset with a different environment. If the goal is to discard environmental cues and focus only on the pedestrian interactions, the trajectories should be suitably normalized to present a uniform point of view. In Fig. \ref{862218}, the Zara01 and Zara02 datasets obviously share the same environment, while the UCY dataset has a different "floor plan": using image coordinates would work well for the first two but clash with the third dataset.