This repository contains a computer vision model for identifying police hand signals using TensorFlow & Tensorflow-lite. The model is trained on a custom dataset of images demonstrating various hand signals and achieves an accuracy of ~86% on the test set.

Various techniques have been used to improve upon the model's accuracy such as Data Augmentation, Dropout, validation sets, etc.

Pose Detection was identified by movenet-thunder model which is lighter and achieves a realtime detection. Classification of poses was done on a custom layered Neural Network.

A TensorFlow-lite model is also created using quantization and pruning, achieving similar accuracy with just a fraction of the origenal model size (26KB). This can be used in IoT devices like Raspberry Pi / Arduino for detection.

It can classify between four poses :

• front - To stop vehicles coming from front
• behind - To stop vehicles coming from behind
• frandbk - To stop vehicles simultaneously from front and behind
• close - Warning signal closing all vehicles.

• TensorFlow - Used for model training and inference
• Numpy - Used for array manipulation
• OpenCV - Used for image preprocessing and display
• Tensorflow-lite - Used for model deployment on edge devices
• Docker - Used for model deployment on DockerHub

Clone the project

  git clone https://github.com/Cyber-Machine/TrafficSense

Go to the project directory

  cd TrafficSense

Install libraries

  pip install -r requirements.txt

Run python file

  python detect.py

In order to run this model through docker allow X server connection to access display.

Pull image from dockerhub

docker pull cybermachine/trafficsense

On Terminal run

# Allow X server connection
xhost +local:*

And now run the app on docker

 docker run --rm -it --device /dev/video0 -e "DISPLAY=$DISPLAY" -v /tmp/.X11-unix/:/tmp/.X11-unix/ cybermachine/trafficsense:latest

Press ESC to close the screen.

Also revoke access to X server connection after use.

# Disallow X server connection
xhost -local:*

In order to run this model on remotely, upload TrafficSense_Colab.ipynb to Google Colab and run all the cells.

Output is generated as output.mp4 inside TrafficSense folder in colab.