Object Detection

Object Tracking (OT)

Norfair recommended by Mahesh for 2D, combines with YOLOv5 / YOLOv7 Given a sequence of detections (from multiple sources), uniquely identify independent actors in the scene and track their movement across time.

We need to do Association with Object Tracking (essentially have a bunch of Object Detection, and then relate those bounding boxes).

Just use a Kalman Filter? https://towardsdatascience.com/what-i-was-missing-while-using-the-kalman-filter-for-object-tracking-8e4c29f6b795a

Charles Zhang: Hey Steven, sorry I missed this message. There’s nothing really outstanding on OT - it’s a working simple product that does OT reasonably reliably based on the AB3DMOT tracker.

• Future steps would be to investigate learned approaches to object tracking & fusion to try achieve better tracking performance at the cost of higher compute
• If you have questions I’ll be happy to answer any Otherwise, next steps would be to see if downstream or upstream have any pain points to alleviate (Rowan?)

Association: Two stage.

Concepts

• Hungarian Algorithm

WATonomous

The way we do it WATonomous is without any ML.

We just do association, using AB3DMOT

From the article

https://www.v7labs.com/blog/object-tracking-guide

Papers

• SlowFast
• DeepSort
• MDNet