Skip to content

Latest commit

 

History

History
229 lines (164 loc) · 8.12 KB

README.md

File metadata and controls

229 lines (164 loc) · 8.12 KB

node-moving-things-tracker

node-moving-things-tracker is a javascript implementation of a "tracker by detections" for realtime multiple object tracking (MOT) in node.js / browsers.

Commissioned by moovel lab for Beat the Traffic X and the Open Data Cam project.

Problem

How to track persistently multiple moving things from frame-by-frame object detections inputs? How to assign an unique identifier to frame-by-frame object detection results?

Often object detection framework don't have any memory of their detection results over time e.g. Yolo provides every frame an array of detections results in the form of [[x,y,w,h,confidence,name] ...], note that there isn't any unique ID to indentify the same detected object in future frames.

Detections Input

Raw Yolo detection results

detections

Tracker Output

Yolo detection results enhanced with unique IDs after beeing processed by node-moving-things-tracker. Note that now every object has been assigned an unique ID

tracker

Installation

# Install globaly to use as command line tool
npm install -g node-moving-things-tracker 

# Install localy your node.js / javascript project
npm install --save node-moving-things-tracker 

Usage

As a node module

const Tracker = require('node-moving-things-tracker').Tracker;

Tracker.updateTrackedItemsWithNewFrame(detectionScaledOfThisFrame, currentFrame);

const trackerDataForThisFrame = Tracker.getJSONOfTrackedItems();

Example: Opendatacam.js of the Open Data Cam project.

Command line usage

node-moving-things-tracker takes as an input a txt file generated by node-yolo and outputs a tracker.json file that assigns unique IDs to the YOLO detections bbox.

The detections entry file could also be generated by another object detection algorithm than YOLO, it just needs to respects the same format.

NOTE : usage is customized for the use case of Beat the Traffic X

node-moving-things-tracker --input PATH_TO_YOLO_DETECTIONS.txt
# This will output a tracker.json file in the same folder containing the tracker data

Detections Input and tracker output format

See example here:

Coordinate space:

coordinates of tracker input

Detections Input

rawdetections.txt

{"frame":0,"detections":[{"x":699,"y":99,"w":32,"h":19,"confidence":34,"name":"car"},{"x":285,"y":170,"w":40,"h":32,"confidence":26,"name":"car"},{"x":259,"y":178,"w":75,"h":46,"confidence":42,"name":"car"},{"x":39,"y":222,"w":91,"h":52,"confidence":61,"name":"car"},{"x":148,"y":199,"w":123,"h":55,"confidence":53,"name":"car"}]}
{"frame":1,"detections":[{"x":699,"y":99,"w":32,"h":19,"confidence":31,"name":"car"},{"x":694,"y":116,"w":34,"h":23,"confidence":25,"name":"car"},{"x":285,"y":170,"w":40,"h":32,"confidence":27,"name":"car"},{"x":259,"y":178,"w":75,"h":46,"confidence":42,"name":"car"},{"x":39,"y":222,"w":91,"h":52,"confidence":61,"name":"car"},{"x":148,"y":199,"w":123,"h":55,"confidence":52,"name":"car"}]}

Tracker Output

Normal mode:

{
  // Tracker data for each frame
  "43": [
    {      
      "id": 0,
      "x": 628,
      "y": 144,
      "w": 48,
      "h": 29,
      "confidence": 80,
      "name": "car",
      "isZombie": false
    },
    {
      "id": 1,
      "x": 620,
      "y": 154,
      "w": 50,
      "h": 35,
      "confidence": 80,
      "name": "car",
      "isZombie": true
    }
  ]
}

Debug mode:

#Run with ---debug flag at the end
node-moving-things-tracker --debug --input PATH_TO_YOLO_DETECTIONS.txt
{
  // Tracker data for each frame
  "43": [
    {
      "id": "900e36a2-cbc7-427c-83a9-819d072391f0",
      "idDisplay": 0,
      "x": 628,
      "y": 144,
      "w": 48,
      "h": 29,
      "confidence": 80,
      "name": "car",
      "isZombie": false,
      "zombieOpacity": 1,
      "appearFrame": 35,
      "disappearFrame": null
    },
    {
      "id": "38939c38-c977-40a9-ad6a-3bb916c37fa1",
      "idDisplay": 1,
      "x": 620,
      "y": 154,
      "w": 50,
      "h": 35,
      "confidence": 80,
      "name": "car",
      "isZombie": false,
      "zombieOpacity": 1,
      "appearFrame": 43,
      "disappearFrame": null
    }
  ]
}

Run on opendatacam/darknet detection data

Usage with opendatacam/darknet (opendatacam/darknet#2) generated tracker data

node-moving-things-tracker --mode opendatacam-darknet --input detectionsFromDarknet.json
# This will output a tracker.json file in the same folder containing the tracker data
# Output format is the same as previously

Example detections.json file

[
  {
    "frame_id":0, 
    "objects": [ 
      {"class_id":5, "name":"bus", "relative_coordinates":{"center_x":0.394363, "center_y":0.277938, "width":0.032596, "height":0.106158}, "confidence":0.414157}, 
      {"class_id":5, "name":"bus", "relative_coordinates":{"center_x":0.363555, "center_y":0.285264, "width":0.062474, "height":0.133008}, "confidence":0.402736}
    ] 
  },
  {
    "frame_id":0, 
    "objects": [ 
      {"class_id":5, "name":"bus", "relative_coordinates":{"center_x":0.394363, "center_y":0.277938, "width":0.032596, "height":0.106158}, "confidence":0.414157}, 
      {"class_id":5, "name":"bus", "relative_coordinates":{"center_x":0.363555, "center_y":0.285264, "width":0.062474, "height":0.133008}, "confidence":0.402736}
    ] 
  }
]

Run on MOT Challenge dataset

node-moving-things-tracker --mode motchallenge --input PATH_TO_MOT_DETECTIONS.txt

# Output will be in the same folder as input under the name outputTrackerMOT.txt

Benchmark

How to benchmark against MOT Challenge : https://github.com/opendatacam/node-moving-things-tracker/blob/master/documentation/BENCHMARK.md

Limitations

No params tweaking is possible via command-line for now, it is currently optimized for tracking cars in traffic videos.

How does it work

Based on V-IOU tracker: https://github.com/bochinski/iou-tracker/ , paper: http://elvera.nue.tu-berlin.de/files/1547Bochinski2018.pdf

In order to define if an object is the same from one frame to another, we compare the overlapping areas between the two detections between the frames.

screen shot 2017-10-18 at 15 57 25

By computing the intersection over union:

screen shot 2017-10-18 at 16 02 12

On top of this we added some prediction mecanism for next frame based on velocity / acceleration vector to avoid ID reassignment when the object is missed only for a few frames.

License

MIT License

Acknowledgments

node-moving-things-tracker is based on the ideas and work of the following people. References are listed chronologicaly how we encounter them.