Computer Vision Pipeline•CASE STUDY N003

VisionTrack CV

AI Multi-Face Attendance & Mask Analytics Suite

PythonOpenCVYOLOv8Dlib 128-DCustomTkinterSQLite WALSMTP Pipeline

01. OBJECTIVE & SYSTEM CONSTRAINTS

The Engineering Problem

A production-grade computer vision suite engineered for automated multi-face batch recognition, mask compliance scoring, and thread-safe SQLite WAL attendance logging.

Systemic Bounds & Operational Constraints:

▶Consistent 60 FPS GUI responsiveness under continuous live video stream processing.
▶Zero database corruption during high-frequency concurrent worker thread writes.
▶Real-time face vector matching against serialized 128-dimensional embedding stores without blocking the main render loop.

02. STACK RATIONALE & ARCHITECTURE DECISIONS

STACK_RATIONALE // VISION_INFRASTRUCTURE

⚡Multithreaded Worker Isolation(vs. Main UI Thread Execution)

Decouples video frame capture and YOLOv8/Dlib model inference into background worker threads communicating via thread queues, maintaining a smooth 60 FPS GUI experience.

⚡YOLOv8 + Dlib 128-D Vectors(vs. Single Object Detectors / Haar Cascades)

Combines Ultralytics YOLOv8 for mask compliance bounding boxes with Dlib 128-dimensional facial vector matching for high accuracy.

⚡SQLite Write-Ahead Logging (WAL)(vs. Standard Rollback Journaling)

Uses WAL mode with explicit mutex locking (threading.Lock) to guarantee database integrity across concurrent background worker threads.

⚡Automated SMTP Daemon Loop(vs. Manual Absentee Reporting)

Background thread evaluates daily cutoff timestamps (e.g. 09:00 AM) and dispatches automated absentee email notices to parents/students.

03. SYSTEM ARCHITECTURE TOPOLOGY

HOVER NODES TO INSPECT SERVICE BOUNDARIES AND IMPLEMENTATION DETAILS

VISIONTRACK CV PIPELINE // C4_CONTAINER_DIAGRAM

System Boundary: VISIONTRACK CV PIPELINE

// INPUT TIER

RTSP & USB Video Feed

[Frame Acquisition]

OpenCV VideoCapture / RTSP

OPENCV FRAME BUFFER

// PROCESSING TIER

Frame Pre-Processor

[Image Pipeline]

OpenCV / NumPy

YOLOv8 & Dlib Vector Engine

[Inference Engine]

YOLOv8 / Dlib / OpenCV

THREAD-SAFE QUEUE / WAL MUTEX WRITES

// PRESENTATION & PERSISTENCE TIER

CustomTkinter Desktop UI

[Desktop Application]

CustomTkinter / Python

Thread-Safe SQLite WAL DB

[Persistence Store]

SQLite WAL / Thread Lock

RTSP & USB Video Feed [Frame Acquisition]OpenCV VideoCapture / RTSP

Captures live video streams from USB webcams and IP RTSP camera streams.

04. SYSTEM INTERFACES & PLATFORM GALLERY

Operations Dashboard

Student Registration & Dataset Staging

Live Attendance Session & Camera Feed

System Settings & Cutoff Configuration

05. SUBSYSTEM ARCHITECTURE DEEP-DIVE

4 MASTER SUBSYSTEMS — 3 SQLITE TABLES — 8 CORE PYTHON MODULES

CustomTkinter Desktop Suite

GUI ENGINE

Modern dual-theme presentation suite (Dashboard, Register, Attendance, Records, Settings) running background worker threads for 60 FPS UI responsiveness.

0 SQLite tables|5 views

YOLOv8 & Dlib Vector Pipeline

CV PIPELINE

Batch multi-face detection, Ultralytics YOLOv8 mask compliance bounding boxes (Mask, No Mask, Incorrect), and 128-d Dlib face embedding vector matching.

0 SQLite tables|2 workers

Thread-Safe SQLite WAL Storage

SQLITE WAL

Write-Ahead Logging database engine with explicit threading.Lock mutex guards across students, attendance logs, and audit records.

3 SQLite tables|3 tables

Automated SMTP Absentee Daemon

SMTP DAEMON

Background thread daemon scanning arrival timestamps against daily cutoff settings (e.g. 09:00 AM) and dispatching automated email notices.

0 SQLite tables|1 daemon

06. IMPLEMENTATION CHALLENGES & HARD BUGS

CHALLENGE 01

Thread-Safe SQLite Write-Ahead Logging

Problem Encountered:

Concurrent writes from background face detection threads and GUI operations caused database lock contention and crashes.

Architectural Solution:

Configured SQLite Write-Ahead Logging (WAL) mode paired with explicit Python threading.Lock mutex guards.

System Guardrail:Mutex-protected DB transactions guarantee zero database corruption under continuous multi-thread log dispatches.

CHALLENGE 02

Multi-Face Real-Time Batch Detection

Problem Encountered:

Extracting 128-d facial embeddings for multiple students in a single frame dropped video FPS dramatically.

Architectural Solution:

Engineered asynchronous worker threads and a single-item queue buffer system to detect, embed, and identify multiple faces simultaneously without main thread blocking.

System Guardrail:Single-item frame queue buffer prevents memory backpressure during heavy multi-face frames.

CHALLENGE 03

Threaded Desktop GUI Architecture

Problem Encountered:

Running computer vision models on the main UI thread caused interface freezing and unresponsive controls.

Architectural Solution:

Isolated video capture and model execution inside background worker threads communicating via thread queues.

System Guardrail:Complete decoupling of UI render loop from inference execution guarantees 60 FPS GUI responsiveness.

07. TELEMETRY & METRIC STACK

60 FPS

GUI RESPONSIVENESS

Multithreaded background worker isolation

128-D

FACE EMBEDDINGS

Dlib facial feature vector extraction

SINGLE

FRAME QUEUE

Thread-isolated single-item buffer prevents UI lag

100% WAL

SQLITE INTEGRITY

Write-Ahead Logging mode with mutex DB lock

08. ARCHITECTURAL TRADEOFFS & INSIGHTS

⚠ SINGLE-STATION SQLITE WRITE LOCKING: Operating an SQLite WAL database on a single desktop station provides zero-latency local writes, but multi-gate deployment introduces write-lock contention across concurrent stations. Decoupling local frame processing from database persistence via event queues maintains steady 60 FPS GUI performance.

Technology Stack:

PythonOpenCVYOLOv8Dlib 128-DCustomTkinterSQLite WALSMTP Pipeline

← BACK TO ALL PROJECTS