SORT
last modified : 30-03-2020
General Information
- Title: Simple Online And Realtime Tracking
- Authors: Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, Ben Upcroft
- Link: article
- Date of first submission: 2 Feb 2016
- Implementations:
Brief
In this paper the authors tackle real-time object tracking. According to the authors, most of the existing approaches are either fast or accurate. This makes the accurate trackers not usable for real-time applications.
The authors propose a simple approach that does not account for corner cases but works extremely fast. The method is called SORT and is based on Kalman Filter and Hungarian algorithm. SORT also uses a deep detector, which seems to greatly improve the accuracy. The proposed approach achieves state-of-the-art accuracy while being able to update the tracker at a rate of 260 Hz (over 20x faster than other state-of-the-art methods).
How Does It Work
The method is based on a four steps pipeline:
- First, all the objects are detected in the image
- Second, existing tracks positions are updated using a Kalman filter
- Then, all the updated positions are matched to the detections using an Hungarian algorithm
- Finally, un-matched detection are set as new tracks
There are also some rules implemented to remove lost tracks and ignore noisy detections.
Results
They evaluate on MOT benchmark sequences. Hereafter is the graph detailing speed vs accuracy, they provide with a detailled table but do not present the update rate in it, check the paper for more details.
