Nirikshak

Nirikshak: Transforming CCTV Networks into Proactive Crime Prevention Tools

Overview

With over a billion CCTV cameras deployed globally as of 2021, research shows no significant reduction in crime. Cities like Delhi, despite having a high density of surveillance cameras, continue to see alarming crime rates, especially against women—with over 14,000 incidents in 2022 alone.

The root problem? Over 95% of CCTV footage is never actively monitored. Surveillance systems are mainly used for post-incident analysis rather than real-time intervention. Human fatigue and oversight also hinder proactive monitoring, which inspired us to create Nirikshak – an AI-powered application that turns passive surveillance into proactive crime prevention, supporting UN Sustainable Development Goal 5 for gender equality.

What It Does

Nirikshak leverages AI for real-time analysis of CCTV feeds, offering:

Real-Time Monitoring: Deep learning detects suspicious activity in real-time.
Immediate Alerts: Notifies security personnel and authorities with live video, location, and incident details.
Contextual Understanding: Context-aware AI differentiates between types of crimes.
Crowdsourced Reporting: Anonymously report incidents through the app.
Data Analytics: Provides charts and insights on crime patterns to support strategic planning.
Scalability: Capable of managing large CCTV networks in metropolitan areas.

How We Built It

Our solution integrates multiple advanced technologies:

Dataset Creation: Curated a dataset covering crime types like assault, harassment, kidnapping, and vandalism.
Feature Extraction: Utilized Inflated 3D ConvNets (I3D) for feature extraction from video data.
Temporal Context Aggregation: Developed a transformer-based model for learning temporal information from videos.
Prompt-Enhanced Learning: Leveraged Gemini and Gemma models for concept extraction and enhanced context awareness.
Model Optimization: Designed a lightweight model with only 1.21 million parameters, requiring just 241 million operations per second.
Application Development: Created an Android app with modes for surveillance, security personnel, and public reporting.

Challenges We Faced

Dataset Compilation: Collecting and annotating specific crime data was time-intensive.
Model Integration: Compatibility challenges integrating Gemini and Gemma models.
Real-Time Processing: Ensuring the model processes video feeds with minimal latency.
Resource Constraints: Optimizing the model for deployment on existing infrastructure.

Accomplishments We're Proud Of

Context-Aware Detection: Implemented a context-driven system for accurate crime detection.
Model Efficiency: Achieved a lightweight, computationally efficient model.
Deployable Application: Developed an Android app ready for real-world use.
Community Engagement: Integrated crowdsourced reporting to enhance community involvement.

What We Learned

Context Matters: Understanding context significantly enhances detection accuracy.
Integration Complexities: Overcame challenges integrating multiple AI models.
User-Centric Design: Developed features for diverse user groups like security personnel and the public.
Scalability Considerations: Optimized the solution for metropolitan scalability.

What's Next

Pilot Deployment: Collaborate with local authorities for pilot testing.
Model Enhancement: Further refine models for improved accuracy and processing speed.
Expanded Crime Categories: Include more crime types and adapt to different cultural contexts.
Cross-Platform Development: Expand the app to iOS and web platforms.
International Expansion: Deploy Nirikshak globally to enhance public safety, particularly for women and girls.

By transforming passive surveillance into proactive intervention, Nirikshak aims to make public spaces safer for everyone, especially women and girls.