I'm trying to inference people from upwards and count them using Yolov5. I know the controversy between yolov5 and yolov4, but for me, Yolov5 is more easier and reliable to use, also the setup.
I have tried SORT, Deep-SORT to track count people passing the gate, but it gets lost when the camera directly films from the upper side, so the tracking gets lost and maps with new id. When I remove the tracker I can see that the inferencing is not happening.
I'm not sure I can solve this by touching the script, or the only way is to train the model. Currently I have been using pretrained model, which I believe is been trained by COCO dataset. I have been using the people class only from the model.
If I have to train the model, can someone give me guidance about how to do that? I am wondering how to train, as the pretrained model is containing 80 classes, is it better to wipe out everything and train it from scratch, or to use Transfer-learning tool. And if I decide to use Tlt, should I keep the 80 classes, or only left people class? I'm using the people class only, but wiping out the other classes would affect the whole network, from my understanding.
Thanks in advance!