@tdurand went over the readme and tweaked things a little ... and also added myself to the contributors (hope that is okay) etc.

Open:

• extend the Acknowledgments section with your finds
• resolve the TODO .txt vs .json input as noted in the text
• check whether npm publish is correct (it seems there is also the repo under the old name)

Other wise ready for me

Fix for the memory leak.

I've put together a little doc and published the npm module here: http://npmjs.com/node-tracker-by-detections , we could move it later under "moovel" organization.

Please review: https://github.com/tdurand/node-tracker-by-detections

Still 1-2 days of work to get it finished in february.

I have to confirm with a test suite, but I believe that the tracker does not differentiate merges trajectories without checking for object types.

So if e.g. a Person walks in front of a car, they will not be tracked as two separate objects but as one single object, even if the Modell is capable of detecting the objects separately.

with a github action: https://github.com/marketplace/actions/npm-publish

suggestion from @vsaw

See https://github.com/opendatacam/node-moving-things-tracker/runs/5144461748?check_suite_focus=true

Comes from merging #21

The IOU tracker published in december 2018 an extension of their paper that seems to use some of the improvement we already included here: predicting missing frame to avoid id reassignment.. but maybe worth a deeper look to find some new ideas.

Extending IOU Based Multi-Object Tracking by Visual Information (december 2018) : http://elvera.nue.tu-berlin.de/files/1547Bochinski2018.pdf

• use IOU area overlap (for higher frame rate)
• uses predictions (for lower frame rate)

Make params configurables ... , see this research: opendatacam/opendatacam#3

Looking at the NPM semantic versioning guidelines the recent contributions by @akretz will be released as a bugfix under version 0.8.1

I will prepare the updated package.json and release a new version to NPM once @tdurand has given me access to NPM.

Is there any existent code to produce a video output that shows the bounding boxes and the id of each box?

Looking at the following Darknet JSON Stream, the one can see that darknet detects two persons close to each other, yet clearly distinct from another. The tracker however merges both objects into one, thus erasing one object from the scene.

{
 "frame_id":1137132, 
 "objects": [ 
  {"class_id":0, "name":"person", "relative_coordinates":{"center_x":0.447100, "center_y":0.315462, "width":0.033928, "height":0.115056}, "confidence":0.670621}, 
  {"class_id":0, "name":"person", "relative_coordinates":{"center_x":0.472682, "center_y":0.296385, "width":0.028619, "height":0.118678}, "confidence":0.289695}
 ] 
}, 
{
 "frame_id":1137133, 
 "objects": [ 
  {"class_id":0, "name":"person", "relative_coordinates":{"center_x":0.447339, "center_y":0.313765, "width":0.035581, "height":0.126710}, "confidence":0.582253}, 
  {"class_id":0, "name":"person", "relative_coordinates":{"center_x":0.471141, "center_y":0.292986, "width":0.028312, "height":0.126107}, "confidence":0.308533}
 ] 
}, 
{
 "frame_id":1137134, 
 "objects": [ 
  {"class_id":0, "name":"person", "relative_coordinates":{"center_x":0.452491, "center_y":0.302588, "width":0.030738, "height":0.136299}, "confidence":0.546889}, 
  {"class_id":0, "name":"person", "relative_coordinates":{"center_x":0.469706, "center_y":0.294295, "width":0.027532, "height":0.124118}, "confidence":0.533913}
 ] 
}

Hello, 👋 an interesting repository you have here. Could you please also include instructions on how to train custom classifier? 🙂

We currently still test on node 10 which is out maintenance for quite some while. Will update the official support only for node 12.x, 14.x and 16.x which are currently supported LTS releases.

Copy of opendatacam/opendatacam#410

Seems a very good lib if one day opendatacam wants to get rid of node-moving-things-tracker

https://github.com/tryolabs/norfair

Can take care of this ... we could take one of the BTT videos and play it back in a p5.js sketch (left side: video + raw yolo detection results, right side: video + yolo detection results filtered with moving-things-tracker).

@tdurand Do you have some good ugly raw yolo detection results? (tbd in our call this afternoon)

By default we store all the tracked item in a map that grows indefinitely.. Maybe an issue with tracking things for several hours..

This should be optional so we keep memory usage very low.

Need to add a condition to the call to mapOfAllItemsTracked.set() : https://github.com/opendatacam/node-moving-things-tracker/blob/master/tracker.js#L205

Using: https://github.com/felixSchl/neodoc

Main contributions are

• #21

See opendatacam/opendatacam#263 (comment)

how about adding position stabilization while already having a list of frames?
I'd think that'd be a great additional feature for a tracker, wouldn't it?

as seen in https://github.com/tawnkramer/CarND-Vehicle-Detection#3-stablization

Hello there!

Do you have any demo project or guide on how to use node-moving-things-tracker in the browser?

Hi @tdurand. Do you mind if we add to package.json something like this?

"bundle": "browserify main.js --standalone tracker -o dist/moving-things-tracker.js",
"dist": "npm run bundle && browserify main.js --standalone gd | uglifyjs > dist/moving-things-tracker.min.js"

That way I believe people could use it out of the box in the browser via tracker.Tracker.updateTrackedItemsWithNewFrame ... which would be great as e.g. the create coding audience might not always be into node.js.