Track pedestrians from video or camera using OpenCV.
It detects people bounding boxes using a custom yolov4-tiny model and tracks them.
The initial model was the first 29 layers of Yolov4-tiny pre-trained on the COCO dataset on 80 classes.
Then it was trained / fine-tuned just to detect people with a 416x416 resolution for 6000 iterations on the CrowdHuman dataset improving the mean Average Precision (mAP) from 34.82% to 52.41%, and further to 61% increasing the resolution to 608x608 for inference.
To every contributor / researcher that has made possible the following: