Fig. 1: The above framework image shows (a) the differences between the trajectories of existing datasets A2A and the novel dataset A2E, (b-c) the models trained and tested on combinations of A2A and A2E, (d) the proposed set of metrics for evaluating the accuracy, realism, and decidability of models, and (e) a greedy method for selecting the prediction most realistic movement.
Existing human trajectory datasets have limitations in the sense of embodying interactions. They either do not contain Agent-to-Environment (A2E) interactions, or exhibit limited Agent-to-Agent (A2A) interactions at small scale in simple environments. We speculate that many self-centered pedestrians are prone to avoid or mitigate, consciously or unconsciously, the influence of the environments and other pedestrians during their navigation.
In this work, we are proposing datasets that augment A2E and A2A interactions, which may bring benefits for enhancing learning models by encoding more complex dynamics in trajectories.
We propose a comprehensive trajectory prediction dataset A-to-X that consists of a representative set of trajectories, which will enable better generalization under realistic circumstances that are either complex or unsafe and out-of-distribution (OOD) with respect to current datasets.
This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License and is intended for non-commercial academic use. If you are interested in using the dataset for commercial purposes please contact us.
We take your privacy seriously: if you have any concerns about any of the content provided here, please contact us.
