A question regarding a livestock multi-object tracking project in South Korea.

[Ruby-Aero]

Hello, I am an AI developer living in South Korea.

I've recently started a project to track Korean native cattle (Hanwoo) inside a barn, but I'm facing a lot of difficulties. I'd like to ask for help from mikel-brostrom and other developers to overcome these challenges.

First, let me briefly explain the existing service from another company.
They've optimized a yolov5s.pt model by compiling it into .json and .params using Apache TVM. This optimized model is then used on a Linux server running Rust to receive real-time RTSP streams from CCTV cameras. The system detects a cow when it assumes a 'mounting' or 'parturition' posture.

That's the existing setup. My task is to expand on this by tracking multiple individual cattles within the barn 24 hours to determine which one assumes a 'mounting' or 'parturition' posture.

While doing my research, a few questions have come up:

1. YOLOv5 is primarily for object detection. Wouldn't it be difficult or impossible to integrate it with an object tracking algorithm?

2. If integration is possible, which tracking algorithm would be effective to combine with the YOLOv5 model? I've studied and found StrongSort or ByteTrack with ReID integration. I would love to hear your thoughts on this.

3. I understand that the YOLOv5 model currently used is already trained. Do I also need to find and train a new dataset for the tracking algorithm that is specific to our domain (Hanwoo)?

4. I'm not familiar with how to train tracking algorithms. If training is necessary, could you please tell me how I should approach it?

I have limited knowledge and am struggling with this project. Any help would be greatly appreciated.