01 — Project
Retail Footfall Analytics
Multi-camera edge CV that turns RTSP feeds into footfall, attendance, demographics, and uniform compliance.
5
RTSP Cameras
30+
FPS Processing
~15%
CPU Load
REST
API Endpoints
Edge
Local Agents
01 — Overview
High level system flow
Local edge agents run OpenVINO inference on RTSP feeds and push structured events to REST APIs.
RTSP Ingest
Detect
Track
Recognition
Analytics
REST API
RTSP Ingest
Detect
Track
Recognition
Analytics
REST API
OpenCV · OpenVINO · ByteTrack · FAISS · FastAPI
02 — Impact
Business impact
The Problem
Manual counting, badge systems, and audits don't scale across hundreds of retail cameras.
The Solution
Local edge agents run OpenVINO inference on RTSP feeds and push structured events to REST APIs.
The Result
Stable multi-camera footfall and attendance with dramatically lower CPU load.
CPU Load per Stream
(from ~60%)
FPS Sustained
across 8 streams
Camera Streams
on Intel CPU
Events to API
structured REST events
03 — Architecture
Edge architecture at a glance
Each component is optimized for low latency, reliability, and deterministic behavior on resource-constrained PCs.
Input
RTSP Cameras
Edge Agent (Per Store)
Frame Reader
(Threaded)
OpenVINO
Inference
Tracker
(Feature Extraction)
Local Storage
(configs, logs, retry queue)
Face Recognition
(FAISS + PyTorch)
Output
Footfall API
Attendance API
Demographics API
Uniform API
Input
Edge Agent (Per Store)
Output
RTSP Cameras
Frame Reader
(Threaded)
OpenVINO
Inference
Tracker
(Feature Extraction)
Local Storage
(configs, logs, retry queue)
Face Recognition
(FAISS + PyTorch)
Footfall API
Attendance API
Demographics API
Uniform API
04 — Challenge
Engineering Story
From hostile CCTV conditions to a production-safe edge pipeline.
01 · Problem
Production CCTV is hostile to naive CV — lighting, angles, and false employee matches.
02 · Naive approach
Run full-precision GPU models and simple nearest-neighbor face match.
03 · Why it failed
Edge PCs lack GPU budget; false biometric matches poison attendance and footfall.
04 · Our approach
OpenVINO INT8 edge inference plus FAISS top-2 margin gates for employee disambiguation.
05 · Result
Stable multi-camera footfall and attendance with dramatically lower CPU load.
05 — Engineering
Tech stack highlights
OpenVINO
Optimized CPU inference
ByteTrack
Robust multi-object tracking
FAISS + Qdrant
Fast & accurate face matching
FastAPI
Async REST APIs at the edge
PyInstaller
Single-file exe deployment
Engineering Notes
Edge cases, deployment, and operational detail.
Occlusion, reconnect loops, operating-hour aware recovery, and anti-double-count cooldowns.
06 — Performance
Performance snapshot
CPU Usage
Per Stream
↓ from ~60%
FPS (Avg)
Across Streams
30+
Latency (P95)
End-to-end
120ms
Uptime
Last 30 Days
99.8%
Interested in how we built it?
Multi-camera edge CV that turns RTSP feeds into footfall, attendance, demographics, and uniform compliance.