SmartGate

Table of Contents

• Features
• Requirements
• Setup
• Usage
• Contact
• Contributing

Utilizing real-time object detection for endangered animal species under the Jetson Nano using TensorRT engine optimization. This detection model will be used to determine the state of the hardware-configured gate.

Features

Features include:

• Real-time object detection to set the state of the physical gate. The rules are set within the config/ folder. The config.json file contains the general rules to set the list of stimuli that triggers the specified state of the gate (more details in Setup).

• Utilizing TensorRT engine optimization to enhance inference performance under the Jetson Nano. The optimized models are saved as .engine files within the models/ folder. Models are obtained from the Marsupial dataset.

• Web server dashboard to view a live feed fetched from the camera as well as manually control state of the door.

Requirements

Installing the requirements should be ran under the Jetson Nano with the Jetpack SDK. For more info on setting this up, please refer to NVIDIA's official guides for your respective Jetson Nano model:

• Jetson Nano 4GB
• Jetson Nano 2GB.

There are two options users can use to set up the requirements.

1. Install the dependencies locally on the Nano.

git clone https://github.com/TheOpenSI/SmartGate.git
cd SmartGate/setup
./setup_requirements.sh

Optional: Generate the systemd service and automatically enable and start the service to run on boot.

cd SmartGate/setup
./systemd_service_setup.sh

2. Build Docker container.

git clone https://github.com/TheOpenSI/SmartGate.git
cd SmartGate/
sudo docker build -t smartgate:latest .

Setup

Ensure the Requirements are satisfied before proceeding.

Users are free to configure the rules to set the behaviour of the gate specified on a list of stimuli. Within config/config.json a base template would be provided with:

{
     "model": {
        "path": "../models/yolov5s.engine",
        "classes": "../models/classes/yolov5s.txt",
        "confidence": 0.5
    },

    "rules": [
        {
            "objects": ["dog"],
            "action": "OPEN"
        },
        {
            "objects": ["cat"],
            "action": "CLOSE"
        }
    ],
    
    "server": {
        "port": 8080
    }
}

Model Configuration

Note that in the configuration file, any path specified can be relative to the location of the configuration file itself.

• The model section defines the settings for the object detection model
  • path specifies the file path to the trained model (in this case, a YOLOv5 model in TensorRT format). If you are working with a .pt file, you will need to convert it to a .engine file. You can use the following repo to convert it, or follow this tutorial
  • classes points to a text file containing the list of object classes the model can detect. It will need to be in the following format index: class (check the files in models/classes for some examples)
  • confidence sets the confidence threshold for object detection (0.5 or 50% in this example).

Rules Configuration

• rules section define how the gate should respond to detected objects

  • objects array represents the list of strings of the objects that should trigger the specified action. This should be referred from your specified classes file (from models/classes)
  • action is the action to take when the specified objects from objects are detected. Can be either OPEN or CLOSE.

• In this example:

  • If dog is detected the gate should open
  • If cat is detected the gate should close
  • If both are detected the gate should close. This should be the default behaviour of the SmartGate when both stimuli are detected

Server Configuration

• server section contains the settings for the web server
  • port is the port number for which the web server would run under. In the example it's set to 8080

Usage

Ensure the Requirements are satisfied before proceeding.

To get started run the following

cd SmartGate/src/main/
python3 live_detection.py

A web server should run in which the camera stream can be viewed from the main dashboard and dedicated controls to manually control the gate. Note: You may need to run this command with sudo privileges.

Contact

For project supports, please contact Carlos C. N. Kuhn.

Contributing

Work in Progress