yolov4-deepsort VS tensorflow-yolov4-tflite

I have a video stream of a hiking trail and am trying to count the number of hikers per day. The camera never moves. At night you can also see headlamps :). Direction not as important as just a count of bodies passing through the image in a given time period. https://drive.google.com/file/d/1BwLzyHhrLafyAj5kYn-AfDVBtBBVQCsO/view. I've tried https://github.com/theAIGuysCode/yolov4-deepsort on the stream but it isn't picking up the people, maybe they're too low res for TF Object Detection?

I am trying to track objects in a sequence of images in order to count them. I was looking around for robust trackers since in my case, the camera moves with respect to the object. I found the DeepSORT tracker online and it seems like the solution to my problem. However, I am not sure of how I could train it for my own custom classes. I am currently looking at this repository and it seems to almost do the things I want, except for the counting part. Can anyone explain to me how I can train the DeepSORT tracker for my own classes? I am already training a YOLOv4 model on these custom classes. As a result, I have collected a labelled dataset for the training and validation purposes, and if I have to use images for the training.

I have tried to implement the YOLOv4 + DeepSort tracking from the AI Guys code presented here, just to get an idea of how to go about this, but from what I understand, I will need to train the DeepSORT tracker for detecting my own classes. I do not know how to do that.

You can try a detection and tracking paradigm for videos - it would be a lot more accurate than detection only without the need re-detect every frame which spends a lot of compute. E.g.: https://github.com/theAIGuysCode/yolov4-deepsort

And totally unecessary!! Next day, I discovered a GitHub project that simply used conda, as below.

This GitHub repo contains the comparison. https://github.com/haroonshakeel/tensorflow-yolov4-tflite#fps-comparison