Motivation

Play tracking at a low frame rate(acceleration).

Challenges

However, the low frame rate may introduce some accuracy problem:

Matching the wrong car

Predicate Error

We lose some part of the tracking routine when enabling a low frame rate.

ganler commented 4 years ago

Solution

Tracking at a low frame rate, then re-visit the intermediate frames for insured predictions.

1st: Tracking in low frame rate

In this stage, the system will apply detection first and then track each car independently(circles are tracked car). But it will also record insured matched pairs(the yellow arrows), and re-visit those segments later.

2nd: Filtering

Use some magic(RNN, NND) to filter tracklets where the query predicate doesn't occur.

3rd: Uncertainty Resolution

Find segments that are not confident in pair matching. And re-visit the intermediate skipped frames recursively.

4th: Refinement: we may lose the entrance and exit of the tracking sequence.

Re-visit skipped frames!

Query Planning

Bound: accuracy
System parameters: Sampling framerate, "closeness" threshold.

Question: How to select the parameters.

Randomly sample a few sampled video segments.
Apply detection & tracking algorithms at the full frame rate for validation.
Apply the sys on the sampled videos.