Here I implemented a car counter according to this competition: https://www.computervision.zone/dsc/. I used the Yolo5+deepsort repo (https://github.com/mikel-brostrom/Yol…), trained on the given video as a dataset, and modified the code a bit for the purpose of the contest (track.py from line 210-233).
The modifications are the following:
I store the detected ID-s in an array, and I append those which was not in the array to count the cars.
Sometimes there are similar cars, so the algorithm detects them the same as previous cars, so it gives the same ID to a different car. If this happens I multiply the ID with 100 to distinguish it from the previous car and I count it as a new one.
If cars slowly disappear from the screen at the top, sometimes the algorithm detects them as new cars, therefore after 10 frames, I only count those cars which are detected in the bottom half of the screen to ensure that I do not recount cars.