Track a specific person/object

[GOBish]

Hi! I am using coral in a robot application and was hoping someone could point me in the right direction. I have a tensor flow model working well, performance is great. I am working on a person following project, where I would like to designate one person to follow. This can be done manually (don’t need facial recognition or anything like that). Then I would like to have the robot “lock on” to that person so that it can follow him or her. Any recommendations for a good way to do this? Thanks!

[brianthelion]

@GOBish Can you clarify which view of the person the robot will follow? If the tracking algorithm is "designated" to a frontal view of the human and meant to track that same view, then any classic tracking algorithm will do a reasonable job assuming minimal occlusion. If, however, you want to designate to the frontal view and then track the target from behind (as would be the case with a following robot) then you'll need something more sophisticated. Please advise.

[GOBish]

Thank for the reply @brianthelion ! Would be tracking the rear view of the person. If it helps, can designate the rear view of the person to track at the outset. Are you talking about tracking algorithms like those in openCV or a tensor flow one on Coral?

[brianthelion]

@GOBish If you can designate the rear view of the person, and the scene is relatively uncluttered -- ie, no crowds of people, no major occlusions -- then OpenCV tracking will do fine to start.

[GOBish]

What tracker are you using?

[Namburger]

@GOBish Hi there, you can try working form this object tracking project form the automl team it may give you a better idea: https://github.com/google/automl-video-ondevice#for-coral-device

[GOBish]

Ok - I am training a tflite model now and converting it for use on the EdgeTPU. I will see how the automl tracking works.