search

ZQPei / deep_sort_pytorch

MOT using deepsort and yolov3 with pytorch
MIT License
2.78k stars 720 forks source link

Cross camera person re-identification #120

Open sampathkethineedi opened 4 years ago

sampathkethineedi commented 4 years ago

Hi, the tracker object is holding information about the person id and the feature vector and this gets updated every frame. I'm trying to share this information among multiple camera feeds to identify a person in a camera view who has already been detected in a previous camera view. Any suggestions on how to move forward?

ZQPei commented 4 years ago

Hi, I think you shall create a database of pedestrians and their features.

sampathkethineedi commented 4 years ago

Hey @ZQPei I am storing the target_id and the features in a pickle file after each frame for now. What I am looking for is for another deep sort tracking running on a different video to use this. Specifically, the tracker._match should use the target_id and feature from the pickle file to associate appearance.

Sorry, this is not an issue as such. I'm just looking for suggestions if this is something you have thought about. Thanks!

petalatuy commented 4 years ago

Hi @sampathkethineedi Did you get any ideas on this?

sampathkethineedi commented 4 years ago

I’m using a different reid specific model to match person features from other cameras in the last 50 frames to assign an ID to a new detection.