Making the Implementation Simple

first step is download the datasets from the links given below and keep them in respective locations as specified

Download the file below and paste it in the main directory http://www.robots.ox.ac.uk/~lav/Research/Projects/2009bbenfold_headpose/Datasets/TownCentre-groundtruth.top

Download the below .avi, and paste it inside 'experimenting-with-sort-master\test' http://www.robots.ox.ac.uk/~lav/Research/Projects/2009bbenfold_headpose/Datasets/TownCentreXVID.avi

Now, i have included a script named frameExtractor.py in 'test' folder to extract the frame from the video.

Run the script 'frameExtractor.py'

Now simply run main.py

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