Skip to content

Real-time Multi-person tracker using YOLO v3 and deep_sort with tensorflow

License

Notifications You must be signed in to change notification settings

Qidian213/deep_sort_yolov3

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

62 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

QQ group: 姿态检测&跟踪 781184396

注意: 本代码创建时间较为久远,软件版本较老,很多模块功能不够高效,功能简单。 推荐使用别的作者实现的Pytorch版本等。

Some excellent related work

1, https://github.com/xingyizhou/CenterTrack

2, https://github.com/phil-bergmann/tracking_wo_bnw

3, https://github.com/Zhongdao/Towards-Realtime-MOT

4, https://github.com/ifzhang/FairMOT

5, https://github.com/pjl1995/CTracker

Introduction

Thanks for these projects, this work now is support tiny_yolo v3 but only for test, if you want to train you can either train a model in darknet or in the second following works. It also can tracks many objects in coco classes, so please note to modify the classes in yolo.py. besides, you also can use camera for testing.

https://github.com/nwojke/deep_sort

https://github.com/qqwweee/keras-yolo3

https://github.com/Qidian213/deep_sort_yolov3

Quick Start

  1. Download YOLOv3 or tiny_yolov3 weights from YOLO website.Then convert the Darknet YOLO model to a Keras model. Or use what i had converted https://drive.google.com/file/d/1uvXFacPnrSMw6ldWTyLLjGLETlEsUvcE/view?usp=sharing (yolo.h5 model file with tf-1.4.0) , put it into model_data folder

  2. Run YOLO_DEEP_SORT with cmd :

    python demo.py
    
  3. (Optional) Convert the Darknet YOLO model to a Keras model by yourself:

 please download the weights at first from yolo website. 
 python convert.py yolov3.cfg yolov3.weights model_data/yolo.h5

Dependencies

The code is compatible with Python 2.7 and 3. The following dependencies are needed to run the tracker:

NumPy
sklean
OpenCV
Pillow

Additionally, feature generation requires TensorFlow-1.4.0.

Training the model

To train the deep association metric model on your datasets you can reference to https://github.com/nwojke/cosine_metric_learning approach which is provided as a separate repository.

Be careful that the code ignores everything but person. For your task do not forget to change :

[deep_sort_yolov3/yolo.py] Lines 100 to 101 :

      if predicted_class != 'person' : 
           continue 

Note

You can use any Detector you like to replace Keras_version YOLO to get bboxes , for it is to slow !

Model file model_data/mars-small128.pb need by deep_sort had convert to tensorflow-1.4.0

Deep sort 程序结构见 “model_data/DeepSORT”,如有错误欢迎联系修改。

Test only

Speed : when only run yolo detection about 11-13 fps , after add deep_sort about 11.5 fps (GTX1060 6G)

Test result video : https://www.bilibili.com/video/av23500163/ generated by this project

About

Real-time Multi-person tracker using YOLO v3 and deep_sort with tensorflow

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages