README_EN.md

[Japanese/English]

MOT-Tracking-by-Detection-Pipeline

Tracking-by-Detection method MOT(Multi Object Tracking) is a framework that separates the processing of Detection and Tracking.

---

---

09.MOT.mp4

---

Requirement

opencv-python 4.5.5.62 or later
onnxruntime 1.10.0     or later
mediapipe 0.8.9.1      or later ※When using MediaPipe
filterpy 1.4.5         or later ※When using motpy
lap 0.4.0              or later ※When using ByteTrack
Cython 0.29.27         or later ※When using ByteTrack
cython_bbox 0.1.3      or later ※When using ByteTrack
rich 11.2.0            or later ※When using Norfair
gdown 4.4.0            or later ※When using YoutuReID
tensorflow 2.8.0       or later ※When using Light Person Detector with tflite

*If the installation of cython_bbox fails on Windows, please try the installation from GitHub (as of 2022/02/16).

pip install -e git+https://github.com/samson-wang/cython_bbox.git#egg=cython-bbox

Usage

Here's how to run the demo.

python main.py

• --device
  Specifying the camera device number
  Default：0
• --movie
  Specify video file *When specified, priority is given to the camera device
  Default：unspecified
• --detector
  Specifying the Object Detection method to use
  Specify one of yolox, efficientdet, ssd, centernet, nanodet, mediapipe_face, mediapipe_hand, light_person_detector
  Default：yolox
• --tracker
  Specifying the Tracking algorithm to use
  Specify one of motpy, bytetrack, mc_bytetrack, norfair, mc_norfair, person_reid, youtureid, sface
  Default：motpy
• --target_id
  Specify the class ID to be tracked
  If you specify more than one, specify them separated by commas. *If None, all are targeted.
  example：--target_id=1
  example：--target_id=1,3
  Default：None
• --use_gpu
  Whether to use GPU
  Default：unspecified

Direcotry

│  main.py
│  test.mp4
├─Detector
│  │  detector.py
│  └─xxxxxxxx
│      │  xxxxxxxx.py
│      │  config.json
│      │  LICENSE
│      └─model
│          xxxxxxxx.onnx
└─Tracker
    │  tracker.py
    └─yyyyyyyy
        │  yyyyyyyy.py
        │  config.json
        │  LICENSE
        └─tracker

In the directory where each model and tracking algorithm are stored,
Includes license terms and config.

Detector

Model name	Source repository	License	Remarks
YOLOX	Megvii-BaseDetection/YOLOX	Apache-2.0	Use the model that has been converted to ONNX in YOLOX-ONNX-TFLite-Sample
EfficientDet	tensorflow/models	Apache-2.0	Use the model that has been converted to ONNX in Object-Detection-API-TensorFlow2ONNX
SSD MobileNet v2 FPNLite	tensorflow/models	Apache-2.0	Use the model that has been converted to ONNX in Object-Detection-API-TensorFlow2ONNX
CenterNet	tensorflow/models	Apache-2.0	Use the model that has been converted to ONNX in Object-Detection-API-TensorFlow2ONNX
NanoDet	RangiLyu/nanodet	Apache-2.0	Use the model that has been converted to ONNX in NanoDet-ONNX-Sample
MediaPipe Face Detection	google/mediapipe	Apache-2.0	Unused eye, nose, mouth and ear key points
MediaPipe Hands	google/mediapipe	Apache-2.0	Calculate and use the circumscribed rectangle from the landmark
Light Person Detector	Person-Detection-using-RaspberryPi-CPU	Apache-2.0	-

Tracker

Algorithm name	Source repository	License	Remarks
motpy (0.0.10)	wmuron/motpy	MIT	-
ByteTrack (2022/01/26)	ifzhang/ByteTrack	MIT	ByteTrack is a tracker for single class If you want to use the multi-class extended version, please specify "mc_bytetrack".
Norfair (0.4.0)	tryolabs/norfair	BSD 3-Clause	Norfair is a single-class tracker If you want to use the multi-class extended version, specify "mc_norfiar"
person-reidentification-retail	openvinotoolkit/open_model_zoo	Apache-2.0	ONNX model is obtained from PINTO0309/PINTO_model_zoo Because it is a human model, please specify the class with the target_id option when using it.
YoutuReID	opencv/opencv_zoo	Apache-2.0	Since it is a human model, please specify the class with the target_id option when using it.
SFace	opencv/opencv_zoo	Apache-2.0	Please specify face detection for Detector Also, SFace should perform processing to correct the angle of the face vertically before inference, but this source does not support it.

Author

Kazuhito Takahashi(https://twitter.com/KzhtTkhs)

License

MOT-Tracking-by-Detection-Pipeline is under MIT License.

The source code of MOT-Tracking-by-Detection-Pipeline itself is MIT License, but
The source code for each algorithm is subject to its own license.
For details, please check the LICENSE file included in each directory.

License(Movie)

The sample video uses the "Pedestrian Crossing in Milan, Italy" from the NHK Creative Library.