TrackingOptimization

The project consists in developing a framework for the systematic validation of tracking improvements.

Methodology and Tasks

1. choosing an optimization approach to optimize tracking parameters based on tracking annotations (ie tracker training)
2. designing a cross-validation framework on a set of annotation videos, training on each video in turn and evaluating the tracking performance on the rest
3. testing the new framework with a few variations of an existing open source tracker (time permitting)

Resources

• Traffic Intelligence project: tracker and tools to load trajectories, annotations, and compute tracking performance (Measurement of Tracking Accuracy) https://bitbucket.org/Nicolas/trafficintelligence/
• NOMAD: derivative-free (black-box) optimization tool https://www.gerad.ca/nomad/Project/Home.html
• Python PythonResources
• Version control software, either Mercurial or Git, and repository on Bitbucket or Github Mercurial
• Cross-validation https://en.wikipedia.org/wiki/Cross-validation_(statistics)
• Annotated dataset: https://www.jpjodoin.com/urbantracker/dataset.html (other to come)
• Previous work:

1. Dariush Ettehadieh's master's thesis http://publications.polymtl.ca/1661/
2. Paper on transferability http://n.saunier.free.fr/saunier/stock/morse16transferability.pdf