Inspiration
India has millions of CCTV cameras, but very few people actually monitor them. This leads to missed incidents, delayed responses, and wasted resources. We wanted to build an AI agent that doesn’t just "watch" video but actually "understands" it and produces reports that help security staff act quickly.
What it does
SafeSurveil AI automatically processes CCTV footage and turns it into actionable intelligence. It can: -Detect people, vehicles, overcrowding, accidents, or unusual behavior.
-Anonymize sensitive data (like faces).
-Generate human-readable summaries of incidents.
-Send real-time alerts and create end-of-day reports.
In short, it converts raw camera feeds into structured safety insights.
How we built it
1.Data Preprocessing – IBM toolkits for frame extraction, anonymization, and video handling.
2.Perception Layer – Vision models for detection, tracking, anomaly recognition.
3.**Agent Workflow – **Autonomous pipeline: Ingestion→Analysis→Tagging→Summarization
4.Generative Summarization – Granite models for natural language event reporting.
5.Prototype – Tested end-to-end pipeline on CCTV datasets.
Challenges we ran into
-Handling large-scale video streams in real time.
-Balancing accuracy vs latency.
-Creating synthetic datasets for rare anomalies (e.g., accidents).
-Edge deployment under compute/bandwidth limits.
Accomplishments that we're proud of
-Built a full autonomous AI agent pipeline.
-Scalable for India’s dense urban environments.
-Produced structured AI intelligence reports (not just detections).
What we learned
-The value of combining computer vision + generative AI.
-The importance of data ethics and anonymization.
-How synthetic rare-event data improves training.
What's next for SafeSurveil AI
-Deploy on edge devices for real-time detection.
-Add multi-camera coordination.
-Partner with government & city authorities for pilots.
-Support multilingual reporting for India.
Log in or sign up for Devpost to join the conversation.