Now that you have a potentially viable re-identification module, the next step is to incorporate it into your system. That involves adding in the following functionality:

1. When a person enters, or soon thereafter, you should instantiate a Guassian mixture model for the person as their identifier. Store it with the persons information.
2. When a person leaves, rather than discard them and their ID, place their model and information into a list of persons who have left the scene.
3. When a person re-enters and has a model instantiated, it should be compared against the existing models of the persons who have left. If the model matching meets a certain matching score threshold, then the person should be considered a match. Given all potential matches, select the one with the best matching score. The ID should be modified to be the original ID of the matched person.

Note: Given that a threshold is needed, you will have to play around with your training and testing sets to see what is a decent matching threshold. Naturally, you may not get 100% re-identification. Try to err on the side of less false positives. It might be better to not re-connect a person who re-entered rather than to re-connect two persons who are different.

Deliverable: Apply the re-identification enhanced surveillance pipeline to two videos (can be a sufficiently long subset of a video if your videos are super long). Provide the person count before and after implementation of re-identification. Provide the actual person count as determined by reviewing the video, or through some other means that will give you this information.

Discussion: Discuss how well it works. Are you able to reduce the person count by a reasonable amount so that it much closer to the true person count?

Back