Open-source project
Smorodov/Multitarget-tracker avatar
Smorodov/Multitarget-tracker

Multitarget-tracker: six background-subtraction detectors documented, and a changelog line that removed some

Multiple Object Tracker, Based on Hungarian algorithm + Kalman filter.

2,425 stars672 forksC++Apache-2.0

At a glance

What is it?
This is a C++ multi-object tracker built on Hungarian matching with Kalman smoothing, with a wide menu of detectors, five visual-search trackers and three threading pipelines. The parts worth reading are a removal notice that does not say what it removed, a Python packaging file that imports modules the standard library no longer ships, and a build instruction list for the Python side that the readme never mentions.
Who is it for?
Multitarget-tracker is a serious piece of computer-vision infrastructure with a documentation page that has clearly been assembled over years rather than written at once. The component lists are unusually complete.
Can I use it commercially?
Yes. Apache-2.0 is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
Is it still maintained?
Yes. The repository received new commits within the last day.
What is it written in?
Mainly C++, according to GitHub's language statistics.

Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.

From daily snapshots

Star history

  • Smorodov/Multitarget-tracker2,425 ★ · +6
Star history
Repository2026-09-302026-11-15
Smorodov/Multitarget-tracker2.4k2.4k

6 Stars in the last 46 days, measured from the daily GitHub snapshot. Compare with other projects

Editorial analysis

Six background-subtraction detectors, and a line saying some were removed

The detector documentation is the strongest part of the page and also the part that contradicts itself. Three methods are labelled built-in, three come from the OpenCV contrib module, and all six are given with the exact enumerator name a caller must pass. Then the change list, in its one cleanup entry, says a big code cleanup removed some background-foreground detectors along with some old trackers, some face and pedestrian detectors and an old backend, without naming a single one of them. So a reader has six documented options and a note that the count was lower, and no way to reconcile the two from the page. The build recipe then sets the option that compiles that whole module to on by default. If you followed the documented install exactly and asked the library for one of the removed methods you would find out at runtime.

The one tagged release is named for the version you do not want

There is exactly one release in this repository and its name is Classic version. It was published at the end of October 2025, while the last push to the repository is in September 2026. Every model family in the feature list, every TensorRT detector and the whole background-subtraction cleanup post-date that tag. So the release is a snapshot of the previous code line, published under a version number that reads like a first release rather than like a legacy branch. Anyone installing by tag gets the classic tracker and none of the current detectors. Anyone who wants what the readme describes has to build from the branch tip, which the readme does tell you to do but never says is a different thing from installing the release. The version history here is one entry long and that entry points away from the interesting work.

Six of the twelve feature items are the same item in different fonts

The change list runs to twelve entries and they are not twelve features. Six of them are the same sentence with a different model family substituted in: export a pre-trained PyTorch model from somebody else's repository to the interchange format, then run this tracker with one example number. The families named in that sentence are numerous and span detection, oriented detection and segmentation variants, and two of the six carry an identical typo for the same repository. What the project actually contributes in those lines is a path in the example registry that accepts a given exported file. That is a real and useful contribution and it is not nothing. But presented as six separate items in a twelve-item list of new state of the art, it makes the change log read as six additions rather than as one mechanism applied six times.

The Python packaging imports modules the standard library dropped

The repository contains a Python package file whose purpose is to build the native libraries and install them as an extension module, and whose first lines import four names from the packaging module that Python removed from the standard library in version 3.12. It also imports a version-comparison class from the same module, deprecated years earlier, and sets a module-level metaclass attribute that only had meaning in Python 2. None of that is a style complaint: it means the file cannot be imported at all on a current interpreter, so the Python route to using this tracker is broken on any machine with a modern Python. The readme's install section never mentions it. The only documented way to build the project is the native path with six CMake switches, and the Python package name that the file declares appears nowhere in the readme.

A maintainer's unanswered question is still sitting in the build file

Two classes in that same packaging file exist only to work around the install tool, and both carry docstrings explaining why. The first says there seems to be no other way to do this, describes several attempts that did not work, and ends with a note that help would be appreciated. The comment is still in the file, in the past tense of someone who has since moved on, and the workaround it documents still depends on the packaging behaviour it could not get around. The second class says it skips building anything because everything is already built, and forwards the compiled libraries to the install step, announcing that it is moving library files. In the visible portion of the file that announcement is followed by nothing else. These are the comments of a real maintainer solving a real problem in public, which is worth more than a tidied file would be. It also tells you that the Python install path is load-bearing in a way nobody has revisited.

The usage guide documents a flag syntax the examples never use

The synopsis block lists long-form flags with an equals sign, and the three filename placeholders in it contain an invisible character between the angle bracket and the word, which is the kind of thing that survives copy-paste and breaks a shell. The single worked example on the page uses the short form instead, and so do the six feature-list sentences that tell you which example number to run.

The one command the page shows:

bash
./MultitargetTracker ../data/atrium.avi -e=1 -o=../data/atrium_motion.avi

So the page has two spellings of every option, the synopsis showing one and every actual command showing the other, with no line reconciling them. A reader who copies the synopsis will guess at the short forms correctly and a reader who copies the example will not find those spellings documented anywhere. Neither the build instructions nor the pipeline section resolves it.

Three typos in the removal notice, one duplicated citation, one missing identifier

The cleanup entry, which is the only place a reader learns that things were taken out, contains three spelling errors in one line: a plural rendered without its final letter, a detector name missing two letters, and an abbreviation that is never explained on the page. The Hungarian assignment algorithm is labelled as Hungarian in the prose and misspelled in the enumerator name that callers pass, which is the one error that is a public API rather than a typo. One paper by the same two authors is cited twice, once as a journal item with a page anchor and once as a bare reference. And of the three processing pipelines, the first two are given enumerator names and the third, the four-thread one, is not given any name at all, even though it is the one recommended for the low-frame-rate detectors the feature list spends most of its lines on.

Editorial conclusion

Multitarget-tracker is a serious piece of computer-vision infrastructure with a documentation page that has clearly been assembled over years rather than written at once. The component lists are unusually complete. Six background-subtraction detectors with their exact enumerator names, two assignment algorithms with their complexities stated, two Kalman variants with their state models, and five correlation-filter trackers for the case where a target is lost. That care is what makes the rough edges worth reporting rather than excusing. Two of them are documentation contradictions a user will hit within an hour of building it. Three things to check before you depend on it. Which background-subtraction detectors actually survive, because the page lists six while the change log says some were removed without naming any, and the build switches that whole module on. What the one tagged release contains, because it is named Classic version and predates every model the feature list advertises. And whether you want the Python surface at all, because the packaging file cannot be imported on a current interpreter and the readme's install section only ever describes the native build.

Frequently asked questions

What does Multitarget-tracker do?

It tracks multiple objects through video by solving an assignment problem between detections and existing tracks, then smoothing each trajectory with a Kalman filter. Detections come either from background subtraction or from a neural detector run through an accelerated backend, and a separate visual-search step re-acquires targets that leave the frame using five correlation-filter trackers.

Which detection methods does Multitarget-tracker support?

Six background-subtraction methods are documented with their enumerator names, three of them built in and three coming from the OpenCV contrib module, plus two neural backends. A separate change entry says a code cleanup removed some background-foreground detectors without naming which, while the build recipe still compiles that module.

How do I build Multitarget-tracker?

Clone, make an empty build directory, run CMake with six options all switched on, covering background subtraction, two tracking modules, the accelerated backend, the asynchronous detector and a counting feature, then build with make. No minimal configuration is documented, and the Python package file in the repository is never mentioned in the install section.

Can I install Multitarget-tracker with pip?

The repository contains a Python packaging file that declares a package name and builds the native libraries as an extension, but the readme's install section covers only the CMake path. The packaging file imports modules that were removed from the Python standard library, so it cannot be imported on a current interpreter without edits.

What are the newest features in Multitarget-tracker?

Twelve change entries, six of which are the same item: take a pre-trained model from another project's repository, export it to the interchange format, and run this tracker with one example number. The families named include several current detection, oriented-detection and segmentation lines. The list also announces a large cleanup that removed older detectors, trackers and an old backend without saying which.

Official sources

  1. Issues
  2. License: Apache-2.0
  3. README
  4. Releases
  5. Smorodov/Multitarget-tracker on GitHub
Add this badge to your README

If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.

Add this badge to your README

markdown
[![Hysen Labs](https://hysenlabs.com/badge/smorodov-multitarget-tracker.svg)](https://hysenlabs.com/projects/smorodov-multitarget-tracker)

Star history chart for your README

markdown
[![Star history](https://hysenlabs.com/chart/smorodov-multitarget-tracker.svg)](https://hysenlabs.com/projects/smorodov-multitarget-tracker)

Was this useful?