Fast maneuver recovery from aerial observation: trajectory clustering and outliers rejection

The implementation of road user models that can reproduce a credible realistic behavior in a multi-agent simulation is still an open problem. We propose a data-driven approach to sift through trajectories from a specific scenario and separate them into macro-maneuvers from raw data. Two clustering methods are proposed, mixing the low complexity of the hierarchical clustering method class of algorithms with a post-processing split-merge mechanism based on the initial and final points of each trajectory. The proposed methods’ goal is to isolate non-representative trajectories in their own clusters and in the process to keep the integrity of the ones that are semantically compatible across scenarios. To evaluate the clustering results over a range of methods, a new metric is proposed, called spread on cluster. Combined with the Davies-Boudin (DB) index, it allows to assess the final results according to the existence of outliers and the excessive number of clusters. Considering pedestrians, cyclists and cars in a range of different scenarios, the results confirm the proposed methods effectiveness in detecting different macro-maneuvers while being resistant to the presence of erroneous trajectories in the input data.