Inspiration

Child bullying has been growing rapidly. In 2015, 40% of child suicides in Indonesia happened because of bullying. Most of them happened in school. On average, roughly 67.9 percent of senior high school students, and 66.1 percent of juniors in big cities in Indonesia claim to have been victims of either verbal or physical abuse. Statistics from Indonesia’s child protection commission indicates that such incidents among children are on the rise – they more than doubled from 2,178 reported cases in 2011 to 5,066 in 2014.

What it does

Djook school monitoring will be installed on an existing surveillance camera in Middle School and High School with a pre-installed camera monitoring system in 5 cities in Indonesia with a high rate of physical bullying. This system will notify the school and the parents of the targeted bully act via Facebook Messenger when one of the surveillance cameras recognized a violent activity.

Privacy

At Djook, we believe privacy is a fundamental human right and so much of the personal information. To respect the school privacy about who, when, and how the bully happened at school it will remains stored on each school local server. The information about the students, bully act and the Djook software itself will be host internally by the school in their system. Under all of the students parents and school consent, Djook will only collect pictures of bully act and its school feedback in order to improve Djook machine learning model.

How I built it

Djook has 2 different interfaces: facebook messenger and the camera monitoring dashboard. The camera monitoring dashboard will take live video on each installed camera to be shown on its web interface build with React. The web will send the sliced frame of its video with a certain amount of time interval to the recognition service to get the context of whether the activity recorded on the camera has any violence activity. The recognition service will take 2 contexts from the image: whether the image is labeled violence and who is being targeted in the activity. If true it has violence the service will broadcast a message to its registered stakeholders based on its recognition context via Facebook Messenger.

Challenges I ran into

  • The difficulties to reproduce violence activity to test the trained machine learning model.
  • Processing video with time interval sliced frame may lose the information of activity that happens in between 2 frames.
  • Finding the specially trained model and/or dataset to detect violence activity.

Accomplishments that I'm proud of

  • We have the system up and running during the hackathon and successfully sending messages on Facebook Messenger using Laptop webcam as the "surveillance camera".

What I learned

  • At first, we tried to train an open source PyTorch model for violence activity using our dataset. This is actually my first time working with PyTorch and Facebook messenger API and I learn what possible and what don't. Then we decided to use computer vision services provided by AWS and GCP (Amazon Rekognition and Google Cloud Vision)

What's next for Djook

  • We will try to find out how to make the video to be optimized analyzed using machine learning.
  • Handling multiple cameras to recognize any violence activity with real-time response.

Built With

Share this project:
×

Updates