Modular AI CCTV Platform
Budget: ₹750 – ₹1,250 INR
I’m putting together a single, modular platform that can ingest live CCTV feeds and, in real time, handle security voice-downs, recognise faces for attendance, and spot both fire and sparks. It must also escalate any incident through a built-in response and management workflow rather than just issuing basic alerts.
The same codebase will need to adapt seamlessly across factories, warehouses, offices, schools, hospitals, residential societies, and similar industrial facilities. That means clean separation of modules so I can enable or disable features as each site demands without rewriting or redeploying the whole stack.
Core modules to build
• Security + public-address engine able to push pre-recorded or TTS warnings through existing PA hardware.
• Face-recognition and attendance tracker that logs entries/exits and flags unknown faces.
• Fire and spark detection vision model tuned for indoor and semi-outdoor environments.
Real-time event management
Every detection should feed a central incident console that lets supervisors verify, label, and close events, while archiving video snippets, metadata, and operator notes for later audit.
Key expectations
• Live video ingestion from standard RTSP/ONVIF streams (4-8 MP typical).
• Accurate detections with minimal false positives; configurable confidence thresholds.
• Simple UI (web or desktop) showing multiview, playback, and incident timeline.
• REST or MQTT hooks so the system can talk to third-party alarms, HR, or maintenance software.
• Docker-based deployment script plus clear documentation so I can roll it out on-prem or to a private cloud.
Acceptance criteria
1. Demonstrate each module on recorded sample feeds, then on two simultaneous live cameras.
2. Trigger at least one incident per module and show it flowing through the response console.
3. Provide full source, build instructions, and admin manual in English.
If any off-the-shelf libraries (e.g., OpenCV, TensorFlow, PyTorch, or YolovX) speed things up, feel free to leverage them—just keep licensing clean.
The same codebase will need to adapt seamlessly across factories, warehouses, offices, schools, hospitals, residential societies, and similar industrial facilities. That means clean separation of modules so I can enable or disable features as each site demands without rewriting or redeploying the whole stack.
Core modules to build
• Security + public-address engine able to push pre-recorded or TTS warnings through existing PA hardware.
• Face-recognition and attendance tracker that logs entries/exits and flags unknown faces.
• Fire and spark detection vision model tuned for indoor and semi-outdoor environments.
Real-time event management
Every detection should feed a central incident console that lets supervisors verify, label, and close events, while archiving video snippets, metadata, and operator notes for later audit.
Key expectations
• Live video ingestion from standard RTSP/ONVIF streams (4-8 MP typical).
• Accurate detections with minimal false positives; configurable confidence thresholds.
• Simple UI (web or desktop) showing multiview, playback, and incident timeline.
• REST or MQTT hooks so the system can talk to third-party alarms, HR, or maintenance software.
• Docker-based deployment script plus clear documentation so I can roll it out on-prem or to a private cloud.
Acceptance criteria
1. Demonstrate each module on recorded sample feeds, then on two simultaneous live cameras.
2. Trigger at least one incident per module and show it flowing through the response console.
3. Provide full source, build instructions, and admin manual in English.
If any off-the-shelf libraries (e.g., OpenCV, TensorFlow, PyTorch, or YolovX) speed things up, feel free to leverage them—just keep licensing clean.