-
-
altered angle
-
blurred lens
-
capa
## Inspiration
The inspiration for CCTVSECURITY-AI came from a real and alarming case shown on the news: a criminal who burglarized condominiums for 10 years without being caught. His strategy was simple, but it exploited a critical flaw in current security systems: upon entering the residences, he simply changed the angle of the cameras.
The result was frustrating for the victims. The homeowners only realized the crime hours later, upon returning from work. When checking the recordings, they saw the face of the thief at the moment he shifted the camera, but the burglary itself happened in the "blind spot." The camera passively recorded its own inaction, without anyone being notified at the exact moment of the violation. We realized that cameras today are "witnesses," but not "agents." The CCTVSECURITY-AI was created to fill this 10-year gap of impunity, transforming the monitoring system into an active sentinel that understands when it is being manipulated and alerts the human the second security is compromised.
## What it does
The CCTVSECURITY-AI is an autonomous agent that monitors the quality and positioning of security cameras in real time. It uses local Computer Vision to detect three critical failures:
- Total Obstruction: Covered lens or paint.
- Tampering/Dirt: Blurred lens or lost focus.
- Angle Vandalism: Physical change in the camera's position.
When an anomaly is detected, the agent triggers a Multimodal AI model in the cloud to perform a technical diagnosis and immediately sends a photo with the verdict to the user's Telegram.
## How we built it
We built the system using a hybrid architecture:
- Local Processing: Python and OpenCV perform initial filtering (Blur, Brightness, and SSIM) to ensure cost efficiency and immediate response.
- AI Brain: We use Amazon Bedrock with the Amazon Nova Lite model, integrated via Strands Agents SDK and Boto3. The AI acts as a "hardware expert" that validates whether the local alert is real.
- Alert Interface: Integration with Telegram Bot API for real-time push notifications.
- Calibration UX: We implemented a "Guided Calibration" flow where the user defines the "Golden Point" (initial position) to avoid false positives.
## Challenges we ran into
We faced significant challenges with geographical restrictions of specific models (Anthropic), which led us to quickly pivot to the Amazon Nova model, proving the versatility of the Bedrock platform. Another technical challenge was creating a sensor hierarchy so that the system wouldn't confuse "putting your hand on the lens" with "changing the camera angle".
## Accomplishments that we're proud of
We are proud to have created a "Intelligent Cooldown" system, where the video remains fluid and in real time, while the AI logic and notifications operate in the background without freezing the interface. We achieved high accuracy in obstruction detection even in varying lighting conditions.
## What we learned
We learned how to orchestrate multimodal AWS models for technical visual inspection tasks. We also deepened our knowledge in optimizing computer vision algorithms to run efficiently on edge computing devices.
What's next for CCTVSECURITY-AI The next step is to expand the agent to support multiple cameras simultaneously and integrate with AWS (S3) cloud storage services to create a historical log of vandalism attempts, as well as support for voice commands via Alexa to ask "What is the status of my cameras?".
Log in or sign up for Devpost to join the conversation.