Tracking humans over long durations (months) #5171
Replies: 3 comments 2 replies
-
Also I see that it actually fails to track "uniquely" humans like a kind of face recognition but for the whole body, I guess tracking uses a simple heuristic? Do I have to train a custom face rec for the body to achieve that? |
Beta Was this translation helpful? Give feedback.
-
Hello @louis030195 , the trackers like Botsort and ByteTrack won't go a long way without configuration of them. You can do so by changing the parameters available in the botsort.yaml and bytetrack.yaml. The point here to note is that when you are tracking human movements over months, it is feasible to generate the output in .csv. The Yolov8 framework will give you txt files of the detections which you can continously convert. Also, with a little modification in the code, you can set the name of the detections as date and time.txt which allows you to have a better csv. Also, to reinforce your model better, you can save_crop all your detections and can pass them through opencv harcascade classifiers and lbph recognizers to create a lightweight recognition package that continously trains on your saved_crops and once you get a hit(recognition), you can mention the same in your csv by editing. Hope this helps. |
Beta Was this translation helpful? Give feedback.
-
Hey do you have any recommendation do track human over long durations, like months?
I'm not super familiar with object tracking, but I assume the tracker compute a kind of embeddings of each object and keep a dictionary in memory, upon detecting new object check if the embedding is close enough to existing object's embedding in memory it will use its ID?
So to do this over long durations I would need to save this dict on the filesystem probably in a persistent storage and use it every time I run the tracking.
Any thoughts? (any code would be super helpful, otherwise will dig by myself) 🙏
Beta Was this translation helpful? Give feedback.
All reactions