BRAC CIN: Intrusion & Perimeter Monitoring
Configurable intrusion analytics supporting polygon occupancy and directional line crossing, with track-aware confirmation, cooldowns, evidence storage, and retryable API delivery.
Module of BRAC CIN; presented as a separate case study.
Workflow at a glance
Detect and track
Track people within configured camera views.
Evaluate boundaries
Check region occupancy or line crossing.
Confirm over time
Apply confirmation, cooldown, and deduplication.
Deliver evidence
Queue snapshots and retry event delivery.
Project facts & reported results
Implementation fact
Implementation fact
Implementation fact
Project Overview
Problem Statement
Simple motion alarms cannot reliably distinguish a policy-defined perimeter breach from lighting changes, passers-by, or brief detector noise.
Approach & Methodology
Built a scalable multi-camera surveillance architecture with a GPU Inference Pool (YOLO11s). Designed dual operational modes: Polygon ROI mode for region monitoring and Directional Line-crossing mode for boundary breaches. Implemented two-stage TrackAlertVoter debouncing for reduced transient alerts.
Outcome & Scope
Translates camera-specific polygons and lines into confirmed, deduplicated security events across multi-camera GPU workers.
Technical Challenges & Solutions
Transient Glitches & Stream Freezes
Network instability caused camera feeds to freeze or drop frames, confusing traditional detectors.
Built EdgeGateway with frame-drop detection, automatic stream reconnection, and temporal voter debouncing.
System Architecture
Multi-GPU video analytics engine with dual-mode intrusion detection
Key Components:
Handles stream capture, FPS throttling, and frozen-stream detection
Distributes video frames across multiple GPUs for high-throughput YOLO inference
Processes Polygon ROI containment and Directional Line-crossing logic
Sliding-window temporal voting ensuring persistent detection before triggering alerts
Key Features & Capabilities
Dual intrusion mode: Polygon ROI and Directional Line-crossing detection
Multi-GPU inference pool scaling across multiple video streams seamlessly
Two-stage TrackAlertVoter debounce pattern to reduce transient false alarms
Edge gateway with FPS throttling and automatic frozen-stream reconnect
Automated alert storage with SQLite & daemon sender for CIN API