Globally Consistent Multi-People Tracking using Motion Patterns

Created by Andrii Maksai at CVLAB, EPFL.

This is an approach for simultaneous tracking and learning the motion patterns of the people on the scene.

This work is released under the MIT License (refer to the LICENSE file for details).

This code accompanies paper "Globally Consistent Multi-People Tracking using Motion Patterns". Link to the arxiv submission to be added soon.

Requirements:

Gurobi, free with academic license. $GUROBI_HOME environment variable should exist and gurobi_cl tool should be available.
g++

To install, make. Run bin\test to ensure that everything should work on your system. It has been tested on some versions of OS X and Ubuntu.

Towncentre dataset from 2D MOT benchmark split into 3 1-minute segments accompanies the submission. Original frames are available from https://motchallenge.net/data/2D_MOT_2015/, and ground truth is available from http://www.robots.ox.ac.uk/~lav/Research/Projects/2009bbenfold_headpose/project.html

To reproduce the results reported in the paper (approximately, as the presented version does not optimize hyperparameters using cross-validation), run bin\eval.
There are 3 possible modes of operation: a: learning patterns from the ground truth, b: improving tracking results of other method using the learned patterns, and c: learning the patterns and improving the tracking in unsupervised fashion without the ground truth.

You can run the examples as follows: bin/ptrack examples/training.cfg bin/ptrack examples/testing.cfg bin/ptrack examples/unsupervised.cfg

Note that you should run training before running testing. The .cfg files in the examples directory explain in comments which entries should be present in .cfg files for each of the 3 tasks. To run it on your data, you need to create your own .cfg file and run bin/ptrack on it.

The tracking results are saved in the MOTChallenge format - see https://motchallenge.net/data/2D_MOT_2015 for more details.
The patterns are saved in the format <Number of points on the pattern centerline N> <Width of the pattern> <X_1> <Y_1> ... <X_N> <Y_N>, or a keyword "NoPattern" for special pattern for anomalous behaviours.
The full result file starts with the number of patterns and the number of trajectories. Then follow the description of all patterns in the format described above. Then follow the description of all trajectories in the following format: <Number of points on the trajectory M> <Number of assigned pattern> <Frame #1> <X_1> <Y_1> ... <Frame #M> <X_M> <Y_M>.

Please contact andrii dot maksai at epfl dot ch for any related queries.

maksay/ptrack_cpp