To count objects in real time, I read that YOLO was usually used. In this blog (https://blog.netcetera.com/object-detection-and-tracking-in-2020-f10fb6ff9af3), it says that SSD is an alternative but there is also somes techniques being used that are too slow for tracking in real time in this blog. Is there anything else ? Can't CNN be used in real time for now ?
For restrictive real-time conditions, I recommend the Kalman filter technique for both background estimation and tracking. It is easy to implement and can run on low-end hardware at 50fps.
The title of your message speaks of "object tracking" but your detailed context talks about object "counting" which is only detection and not really tracking (which involves the evolution of states (position and possibly orientation, etc.) of individual objects).
If you only want to count, i.e. detect objects fast, Haseeb Javed
answer is very good. If you want to track individual objects in time, you may have to look elsewhere or combine YOLOR with an actual tracking algorithm (like Kalman filters) to get better performance and follow the state of individual objects.- Badges
- Science topic