AI-Native Application Development
AI Gateway for Security & Surveillance
We build on-premise AI gateways that convert RTSP/RTMP camera feeds to secure, access-controlled WebRTC streams — with real-time computer vision analytics for anomaly detection and intelligent alerting. Zero cloud dependency.
Gateway Capabilities
- RTSP / RTMP → WebRTC real-time transcoding
- Fully on-premise — zero video data leaves the site
- AI object detection: person, vehicle, intrusion zones
- Anomaly alerting with configurable sensitivity rules
- Role-based camera access control per viewer group
- 100+ concurrent camera streams per node
- Embedded / IoT camera support via C++ WebRTC SDK
How It Works
RTSP/RTMP Ingest & Protocol Conversion
ONVIF cameras, NVRs, and embedded SoCs stream via RTSP/RTMP into the gateway. GStreamer pipelines transcode and packetise each stream for WebRTC egress — maintaining frame integrity and minimising conversion latency.
On-Premise Computer Vision Processing
Access-Controlled WebRTC Distribution & Alerting
Authorised viewers access streams via signed WebRTC URLs. Alert events fire webhooks, SMS, and email with annotated video clips attached. All metadata — events, viewer access logs, alert history — is stored locally in PostgreSQL.
What We Build
Protocol Gateway
RTSP/RTMP/HLS ingest → WebRTC egress via GStreamer. Supports ONVIF cameras, NVRs, and embedded SoCs including NVIDIA Jetson.
Computer Vision Analytics
YOLO v8 for perimeter intrusion, crowd density, fire/smoke detection, and PPE compliance — running on local GPU with configurable sensitivity.
Intelligent Alert Engine
Event-driven alerting with configurable rules, multi-channel notifications (SMS, email, webhook), and annotated alert video clip generation.
Access-Controlled Viewing
Per-camera RBAC. Auth via SSO/LDAP. Signed time-limited WebRTC URLs for secure remote access without exposing stream endpoints.
Distributed Node Architecture
Multi-site deployment with central management plane. Each site operates autonomously — no cloud dependency for core functionality.
Forensic Video Search
AI-indexed footage — query by object type, zone, and time window across weeks of recordings. Results link directly to playback timestamps.
CentEdge vs The Alternative
Cloud-Dependent Video AI Platforms
- Video frames uploaded to cloud for AI processing
- Internet dependency for real-time analytics
- Data sovereignty concerns for government/defence
- Per-camera cloud fees at enterprise scale
- VMS lock-in — hard to integrate with existing systems
- All AI inference runs locally on-site GPU hardware
- Operates fully air-gapped if required
- 100% on-premise — no video data ever leaves the site
- One-time build, no per-camera cloud fees
- Open integration with existing ONVIF VMS systems
Who This Is For
- Manufacturing & Industrial Plants
- Banking Branch Networks
- Government & Defence Facilities
- Hospital & Healthcare Campuses
- Smart City Surveillance
- Logistics & Warehousing
- Critical Infrastructure
- Elevator & Lift Monitoring
Technology Stack
GStreamer
Media Server C++ API
LibWebRTC C++
YOLO v8
OpenCV
ONVIF
Redis Pub/Sub
PostgreSQL
NVIDIA Jetson / A100
Frequently Asked Questions
How many cameras can a single gateway node support?
A single gateway node on NVIDIA A100 hardware can handle 100+ simultaneous RTSP streams with real-time CV analytics running on each. For lighter analytics workloads (e.g., motion detection only), this scales to 200+ streams per node. For deployments requiring more cameras, additional nodes are added to a distributed cluster with a central management plane.
Does the AI analytics require internet connectivity?
No. The entire AI pipeline — YOLO v8 inference, alert logic, WebRTC stream distribution, and local storage — runs on-premise with zero internet dependency. The gateway can operate fully air-gapped. Optional cloud features (offsite backup, remote management) are available but never required for core functionality.
What camera types and protocols are supported?
The gateway supports any camera or NVR that outputs RTSP or RTMP streams, which covers the vast majority of IP cameras from Hikvision, Dahua, Axis, Bosch, Sony, and others. ONVIF compliance is supported for cameras requiring PTZ control and event subscription. For embedded IoT cameras on NVIDIA Jetson or similar SoCs, CentEdge's C++ LibWebRTC SDK is used for direct WebRTC output.
How accurate is the AI object detection?
YOLO v8 achieves 90–95% precision on well-configured cameras in good lighting conditions for person and vehicle detection. For specialised use cases (PPE compliance, smoke detection, crowd density), CentEdge fine-tunes the model on representative footage from your deployment environment. Alert sensitivity thresholds are configurable to balance precision and recall.
How is the forensic video search implemented?
Each frame's AI detection results — object types, bounding boxes, zone assignments, timestamps — are stored in PostgreSQL. A query interface allows operators to search by object type, camera zone, and time range. Results return a list of clips with direct playback links and the detection confidence score. For long-retention deployments, an Elasticsearch index is used for sub-second query performance across months of footage.
GET IN TOUCH
Let’s Build This
Together
Tell us about your project and we’ll return with an architecture overview and engagement proposal within 48 hours.
- hello@centedge.io
- +91 6362 814071
- T-Hub, Hyderabad, India