Apply to my own dataset

[xulw8]

Dr. Jiao,
What a coincidence that I came across your paper and code again. I’d like to apply them to my own dataset, but it’s relatively small. Would that affect the model's performance? Is there a general rule on the dataset size for such cases?

[Yiru-Jiao]

Hi @xulw8,

Thanks for your question. Nice to hear from you again. A smaller size of data indeed could lead to unsatisfactory estimation. We mentioned in the paper that

When the vehicle samples are inadequate to form an intuitively proper driver space, the spacing between vehicles may not be inferred.

In general, we recommend first visualising the position distribution of the vehicles (or road users) surrounding the ego vehicle. If the visualisation shows a driver space, our method should be able to infer it.

[xulw8]

Dr. Jiao, thank you for your reply. If I want to study the conflict between through vehicles and left-turn vehicles now, do I just need to extract their trajectories to better infer the parameters?

[Yiru-Jiao]

Hi @xulw8,

Thanks for this question. If the interactions under consideration are between subject (ego) vehicles and left-turning vehicles, I'm not very sure if a multi-directional driver space can be observed. I hypothesise that by visualising the distribution of surrounding vehicles you'll see a half driver space. In this case, it's still ok to infer the parameters of that half, just you'll need to adjust the algorithm and according code.