YoloV7 | DeepSort and Count

#ComputerVision #RealTimeDetection #RealTimeCounting #DeepLearning #DeepSort #MachineLearning #AI #OpenCV #Python #YOLO #MOT #Docker #DigitalImageProcessing

Principal idea

This project started with the idea of being able to help farm workers who still count cattle by hand.

Starting with this idea, for a project on Digital Image Processing at the National University of Río Cuarto, my research began to achieve the first objective.

As the project developed, I realized that I could further generalize its scope. It wouldn't just be cattle, but could be any class, as long as we can train the detection model.

I have managed to make detections in "real time" (almost 30 FPS) with videos or images from different types of sources.

In addition to the detections, it has been possible to track these objects and detect when they enter a certain region of interest (ROI) defined by the user and make a count by class independently detected, for example:

• person: 2; car: 5.

Information flow

Count algorithm

Results

• One class
  • Source: YouTube
  • Invert Frame: False
  • Classes: car → [2]

Test environment

I’ve done tests with:

• PC: HP-OMEN ce004-la
• GPU: NVIDIA 1060 MAX-Q DESIGN 6GB
• CPU: INTEL I7 7700HQ
• RAM: 16GB
• DISK: SSD M2

Performance

• I’ve obtained almost 30 FPS (detection + tracking time) in the majority of the tests.
• Performance suffers when the number of concurrent detections per frame rises above 10/20. We can suffer a drop in fps (detection + tracking) to about 5 per second.

Problems

• You need a lot of processing capacity.
• Occlusion it is a problem (when the objects are too close)
• When an object get out of the frame we lost the tracking and then the program consider a new object if this come into the frame again.

Ideas for the future

• You can create an API.
• You can re-train the detection model with your own class/classes and run the program.
• You can add an visual interface.
• You can collaborate on this repository integrating new features or solving bugs!

---

Try it yourself. Do a test!

Warnings

• NVIDIA Graphic Card needed to run this project.
• Docker previously installed in your computer.
• NVIDIA Docker image needed.

Configuring your HOST PC

Do this following steps in order.

Install Docker

1. Install Docker

2. Post-Install

Install Nvidia Drivers

• Instalation:

  ubuntu-drivers devices # To know which driver is recommended
  sudo ubuntu-drivers autoinstall # To automatically install the recommended driver
  reboot # We need to reboot the system after the installation

• Checks:

  nvidia-smi # Command to check if the driver was installed correctly: The output must be a list of GPU's and a list processes running on it
  sudo apt install nvtop # Program to check the GPU usage
  nvtop # To run the nvtop

  

Pull nvidia-gpu image

Follow this installation guide.

Pull this repository

git pull https://github.com/aagustinconti/yolov7_counting

Download YOLOv7x pretrained weights

1. Download the pretrained weights of YOLOv7x here.
2. Save them in ./pretrained_weights directory.

Configuring the test

• Into the test_oop.py file you can modify some characteristics of the instance of the class YoloSortCount() before to execute the run() method.
  • By default:
    • Source: WebCamera (0)
    • Invert frame: True
    • Classes: person (ID: 0)
    • Save: False

RUN docker_control.py

To build the docker image if it's not available in your PC and run the container with the propp

python3 docker_control.py

Thanks to

YOLOv7

SORT Algorithm