What the Security Operator Gets
01 • Targeted YOLOX AI Detection
Accurate person and vehicle detection executing locally on the edge. Runs continuously across feeds to isolate human and vehicular activity without latency.
02 • Dynamic Top-Left Prioritization
When an event occurs, the affected camera dynamically jumps to the top-left priority anchor tile on your video wall, ensuring guard attention is never split across a dense grid.
03 • 4-Point Detection Zones (ROI)
Right-click any camera to draw a custom 4-point quadrilateral zone. Restrict alerts to sensitive entryways while ignoring public sidewalks, streets, and swaying trees.
04 • Pre & Post Incident Buffering
Captures 3 seconds of circular pre-alarm footage plus 3 seconds of post-alarm footage. Every incident clip shows how the intruder arrived before the alarm sounded.
05 • Mobile Alerts & Dial-Out Response
Review live streams and bounding boxes on iOS/Android. In the event of a verified intrusion, operators can trigger emergency contacts with a single tap.
06 • Multi-Monitor Video Wall
Spawn independent satellite display windows across secondary monitors or commercial video walls, supporting 1x1, 2x2, 3x3, 4x4, 6x6, and 8x8 layouts.
What the IT Team Gets
01 • Works With Existing Cameras & NVRs
Universal RTSP and ONVIF Profile S/T compatibility. Includes automated WS-Discovery plus a one-click NVR batch importer for up to 64 channels.
02 • Customer-Controlled Video
No centralized video storage cloud. Raw streams, neural inferences, database audit logs, and recordings stay strictly on customer-controlled infrastructure.
03 • Master / Worker Cluster
AiWorkerPool streams raw frames to worker machines in parallel, allowing facilities to distribute compute load across multiple PCs as camera counts scale.
04 • Encrypted Peer-to-Peer Networking
Integrates with Tailscale mesh VPN for secure multi-building or remote worker connectivity without opening inbound firewall ports.
05 • Direct Stream Remuxing
Direct stream remuxing (-c:v copy) writes H.264/H.265 video packets into MP4 files with no CPU/GPU video re-encoding.
06 • Embedded SQLite & Rolling Storage
Embedded SQLite database manages configurations, schedules, and incident metadata locally, while automated storage quotas safely prune oldest files.
Technical Architecture & Optimization
MotionGate EMA Pre-Filter
An Exponential Moving Average filter pre-checks each frame. When scenes are static, AI inference is skipped entirely, saving 20 to 30 ms per frame.
Dual-Stream Decoupled Processing
Substream (low-res) runs AI inference and grid previews to preserve framerate, while Mainstream (high-res) is captured for sharp incident recordings.
Emergency Fail-Safe MP4 Finalizer
If network drops or power is cut during an active incident, an automated recovery routine cleanly finalizes MP4 headers to help preserve recoverable incident clips.
Time-Based Operating Schedules
Set per-camera active hours and days of the week. Restricts heavy inference to after-hours surveillance windows and conserves compute during the day.
Hardware Utilization & VRAM Quotas
Supports CPU-only and GPU mode (CUDA/TensorRT). Explicitly set GPU VRAM quotas and thread worker pools so the system never starves the host OS.
Transport Tuning & Dynamic FPS
Select TCP or UDP per camera to eliminate packet drop and video smearing, with real-time FPS throttling (1–30 FPS) based on motion state.
Tested for 40+ Concurrent Dual Streams
Thanks to lightweight MotionGate pre-filtering, Azrian Guard ignores static empty frames and limits heavy AI processing to true motion events. Tested for 40+ concurrent streams on selected hardware configurations (e.g., modern multi-core x86 with dedicated GPU acceleration). Actual capacity depends on camera resolution, codec, FPS, GPU, and AI configuration.