Another question for the detail of the clustering

[Cocoon-Ray]

Thanks for your reply to the implement explanation of the cluster norm in my last issue. I'm still insterested in this trick:

Due to the presence of boundary examples, real driving trajectories can vary significantly, resulting in a wide range of special trajectories. Using only 20 centroid trajectories to represent these different trajectories may introduce considerable bias in my experiment. It seems that clustering based on L2 distance might not address this issue well. How are these centroids obtained through KMeans? Is there any detailed documentation available? Thank you in advance.

[LegendBC]

You can refer to this script