Object tracking, is TensorFlow the right way to go about it?

Hi there, just need a little advice please. I’m new to ML but have done lots of reading and research about various ways to do object tracking. I want to develop a small system to track the location of a known object in front of the camera. It’s basically a rectangle that moves up and down, forwards and backwards and can tilt. Initially, the object of the exercise is to work out the rectangles position, so I can work out its height from a setpoint. Once get that running I want to extract the angle of it’s tilt and finally extract it’s size so I can tell how far away it is (roughly - between 1 and 6 meters). It’s always the same or similar object that’s being tracked.
Initially I’ll create it for my Android phone to prove the idea, then modify to run on an ESP32 CAM or similar.

Is TensorFlow the right way to achieve this?



I don’t understand if you want to have single object model or not.

For a general bbox tracking pipeline you can see:

Then TF micro could run on that devices.

I own and operate a 24ton digger. I want to track the bucket and use the position to create height info. It seems there are various systems that can achieve the tracking. I have limited ML experience (but very keen to learn) so my question is, Is TF Lite/micro the best way of doing this?


I don’t know about any public dataset on this so you need to collect your own dataset with different holes, weather etc…

You can try to find a good collection of depth estimatation papers in:

An alternative you could try to adapt many 3d human pose estimation network to your escavator pose dataset.


If you can use fiducials you can explore something like:

Nope there won’t be any public datasets… Haha, I doubt anybody has ever linked a digger bucket and TensorFlow together ever… I just happen to live in both worlds (excavators and software) and see a useful application.
Thanks for the info, will plod on and see what happens… :sunglasses: