Object Tracking using YOLOv3, Deep Sort and Tensorflow

This repository implements YOLOv3 and Deep SORT in order to perfrom real-time object tracking. Yolov3 is an algorithm that uses deep convolutional neural networks to perform object detection. We can feed these object detections into Deep SORT (Simple Online and Realtime Tracking with a Deep Association Metric) in order for a real-time object tracker to be created.

Getting started

Conda (Recommended)

# Tensorflow CPU
conda env create -f conda-cpu.yml
conda activate tracker-cpu

# Tensorflow GPU
conda env create -f conda-gpu.yml
conda activate tracker-gpu

Pip

# TensorFlow CPU
pip install -r requirements.txt

# TensorFlow GPU
pip install -r requirements-gpu.txt

Nvidia Driver (For GPU, if you haven't set it up already)

# Ubuntu 18.04
sudo add-apt-repository ppa:graphics-drivers/ppa
sudo apt install nvidia-driver-430
# Windows/Other
https://www.nvidia.com/Download/index.aspx

Downloading official pretrained weights

For Linux: Let's download official yolov3 weights pretrained on COCO dataset.

# yolov3
wget https://pjreddie.com/media/files/yolov3.weights -O weights/yolov3.weights

# yolov3-tiny
wget https://pjreddie.com/media/files/yolov3-tiny.weights -O weights/yolov3-tiny.weights

For Windows: You can download the yolov3 weights by clicking here and yolov3-tiny here then save them to the weights folder.

Using Custom trained weights

Learn How To Train Custom YOLOV3 Weights Here: https://www.youtube.com/watch?v=zJDUhGL26iU

Add your custom weights file to weights folder and your custom .names file into data/labels folder.

Saving your yolov3 weights as a TensorFlow model.

Load the weights using load_weights.py script. This will convert the yolov3 weights into TensorFlow .tf model files!

# yolov3
python load_weights.py

# yolov3-tiny
python load_weights.py --weights ./weights/yolov3-tiny.weights --output ./weights/yolov3-tiny.tf --tiny

# yolov3-custom (add --tiny flag if your custom weights were trained for tiny model)
python load_weights.py --weights ./weights/<YOUR CUSTOM WEIGHTS FILE> --output ./weights/yolov3-custom.tf --num_classes <# CLASSES>

After executing one of the above lines, you should see proper .tf files in your weights folder. You are now ready to run object tracker.

Running the Object Tracker

Now you can run the object tracker for whichever model you have created, pretrained, tiny, or custom.

# yolov3 on video
python object_tracker.py --video ./data/video/test.mp4 --output ./data/video/results.avi

#yolov3 on webcam 
python object_tracker.py --video 0 --output ./data/video/results.avi

#yolov3-tiny 
python object_tracker.py --video ./data/video/test.mp4 --output ./data/video/results.avi --weights ./weights/yolov3-tiny.tf --tiny

#yolov3-custom (add --tiny flag if your custom weights were trained for tiny model)
python object_tracker.py --video ./data/video/test.mp4 --output ./data/video/results.avi --weights ./weights/yolov3-custom.tf --num_classes <# CLASSES> --classes ./data/labels/<YOUR CUSTOM .names FILE>

The output flag saves your object tracker results as an avi file for you to watch back. It is not necessary to have the flag if you don't want to save the resulting video.

There is a test video uploaded in the data/video folder called test.mp4. If you followed all the steps properly with the pretrained coco yolov3.weights model then when your run the object tracker wiht the first command above you should see the following.

Video Example

Webcam Example

This is a demo of running the object tracker using the above command for running the object tracker on your webcam.

Command Line Args Reference

load_weights.py:
  --output: path to output
    (default: './weights/yolov3.tf')
  --[no]tiny: yolov3 or yolov3-tiny
    (default: 'false')
  --weights: path to weights file
    (default: './weights/yolov3.weights')
  --num_classes: number of classes in the model
    (default: '80')
    (an integer)
    
object_tracker.py:
  --classes: path to classes file
    (default: './data/labels/coco.names')
  --video: path to input video (use 0 for webcam)
    (default: './data/video/test.mp4')
  --output: path to output video (remember to set right codec for given format. e.g. XVID for .avi)
    (default: None)
  --output_format: codec used in VideoWriter when saving video to file
    (default: 'XVID)
  --[no]tiny: yolov3 or yolov3-tiny
    (default: 'false')
  --weights: path to weights file
    (default: './weights/yolov3.tf')
  --num_classes: number of classes in the model
    (default: '80')
    (an integer)
  --yolo_max_boxes: maximum number of detections at one time
    (default: '100')
    (an integer)
  --yolo_iou_threshold: iou threshold for how close two boxes can be before they are detected as one box
    (default: 0.5)
    (a float)
  --yolo_score_threshold: score threshold for confidence level in detection for detection to count
    (default: 0.5)
    (a float)

Acknowledgments

• Yolov3 TensorFlow Amazing Implementation
• Deep SORT Repository
• Yolo v3 official paper