Vo, B. N., & Ma, W. K. (2006). The Gaussian mixture probability hypothesis density filter. IEEE Transactions on signal processing, 54(11), 4091-4104.
TABLE I: PSEUDOCODE FOR THE GAUSSIAN MIXTURE PHD FILTER
TABLE II: PRUNING FOR THE GAUSSIAN MIXTURE PHD FILTER
+ The Gaussian mixture PHD filter also suffers from computation problems associated with
the increasing number of Gaussian components as time progresses.
+ A simple pruning procedure can be used to reduce the number of Gaussian components propagated
to the next time step.
Object identity (tracklet) was implemented by the following paper:
Pham, N. T., Huang, W., & Ong, S. H. (2007, December). Maintaining track continuity in GMPHD filter. In 2007 6th International Conference on Information, Communications & Signal Processing (pp. 1-5). IEEE.
There are some differences from the GM-PHD algorithm described in Vo & Ma's paper:
-
I have not implemented "spawning" of new targets from old ones, since I don't need it. It would be straightforward to add it - see the original paper.
-
Weights are adjusted at the end of pruning, so that pruning doesn't affect the total weight allocation.
-
I provide an alternative approach to state-extraction (an alternative to Table 3 in the original paper) which makes use of the integral to decide how many states to extract.
Install packages: opencv-python, numpy, ffmpeg.
Download MOT20 dataset from https://motchallenge.net/data/MOT20/.
Copy MOT20/test/MOT20-04 to the root folder of this source code.
Create output folder inside MOT20-04 to store output track images.
Run: python demo_mot20.py
Tracking Result:
Video: https://www.youtube.com/watch?v=QB7tTMdKpGY&feature=youtu.be
Refer to the original gmphd repository by danstowell