Experimenting with SORT

Experimenting with sort different classical tracking algorithms for realtime multiple object tracking (MOT).

Description:

• This is an experiment on Oxford Town Centre Dataset to compare between kalman filter tracker (a motion model) and dlib correlation tracker (an appearance model) in the domain of realtime tracking of multiple objects (pedestrians) in a video sequence (MOT).
• We used the same data association techniques of sort.

Results:

• Dlib correlation tracker: https://youtu.be/tMuX5TP6uqA
• Kalman tracker: https://youtu.be/SKXk6uB8348

Note:

• The detector/ground truth was used only for ~40% of the time.
• We noticed from the above outputs that Kalman tracker is more robust in highly occluded scenes.
• Kalman tracker is also about 10x faster, and so it is more suitable for realtime MOT.

Dependencies:

• Python 2.7x
• dlib
• scikit-learn
• scikit-image
• FilterPy

Usage:

• To test with dlib tracker (default is kalman):

python main.py --dlib

• To save frames with tracking output:

python main.py --save

• To disable online tracking display:

python main.py --NoDisplay