SCHOOL SECURITY: AUTOMATED SURVEILLANCE AND FACE RECOGNITION FROM CCTV CAMERAS IN SCHOOLS

SCHOOL SECURITY: AUTOMATED SURVEILLANCE AND FACE RECOGNITION FROM CCTV CAMERAS IN SCHOOLS

The Client:

An international school chain with more than 30+ schools.

The Problem:

To ensure the security of children from anti-social individuals, the client was looking for an automated CCTV surveillance system. The client also wanted a Facial Recognition system integrated into the security infrastructure. This was aimed to decrease reliance on CCTV operators and eliminate human errors in judgement, making security management more robust and foolproof.

One of the major challenges was that the CCTV infrastructure consisted of low resolution 1 MP cameras, making it harder to accurately recognise faces.

The Solution:

Using Deep Learning, we built an automated monitoring system for area protection and face detection that analyzed multiple streams in real time and raised alerts in the event of a breach, identifying the violator instantly. The system sends instant notifications using WhatsApp along with 15-second video clips when breaches or violations are detected.

The following workflow was followed to develop and deploy the surveillance system:

  1. Complete infrastructure was designed to get the local camera feeds online so that processing can be done on the cloud and a scalable system can be developed.
  2. Deep learning based gender and age classification models were trained to classify people into student, staff, male and female.
  3. Deep learning based face detection and recognition models were trained.
  4. An intuitive UI was developed to manage the streams, identities, notifications and other meta information about the streams.
  5. WhatsApp integration was provided for sending notifications directly to mobile in case of any breach or violation enabling significantly reduced action time.

Results:

Complete end to end solution was provided which was capable of handling multiple streams and processing them in real time. Our custom trained model was able to recognize faces with 82% accuracy (with 200 unique identities in the target database) despite the low resolution of the cameras.

Below is a screenshot of the dashboard. All streams can be viewed in real time. The dashboard also has a notification panel for managing and viewing all the breaches and violations. Along with this dashboard provides an interface for managing face identities and stream settings, so that new streams and faces can be added on the fly. The dashboard’s ease of navigation allowed rapid adoption of the technology.

Technology Stack:

Python, Tensorflow, Messaging Queues(Kafka, Redis), React, Django

Leave a Reply