← Projects Archive
Field-Proven
Computer Vision & Edge OCR
KVMS
Real-Time Vehicle Analytics & OCR Classification System
98.2%
Plate recognition accuracy
< 120ms
Frame inference speed
24/7
Continuous unattended field testing
Technical Scope
Automating access control and security monitoring requires sub-second vehicle identification without human intervention.
Architecture Highlights
- Attribute & Plate Detection: Two-stage neural network architecture isolating vehicle bounding boxes, extracting plate sub-regions, and parsing character sequences.
- Robustness in Uncontrolled Environments: Pre-processing filters normalizing contrast, skew, and glare under challenging environmental conditions.
System Pipeline Flow Automated Pipeline
STAGE 01
Handheld Video Ingest
5–10s continuous mobile capture sweep
STAGE 02
Frame Extraction & Keyframe Filtering
Photometric sharpness & motion blur screening
STAGE 03
Feature Matching & 3D Point Cloud
Dense multi-view photogrammetric reconstruction
STAGE 04
Deep Learning Semantic Segmentation
Component boundary & defect zone isolation
STAGE 05
Volumetric & Tolerance Calculation
CAD delta metrology to 0.01mm / micron scale
STAGE 06
Automated Reporting & HITL Review
QA dossiers, STL export & validation queue
SQS-Buffered Distributed Ingestion & Dynamic Compute
Decoupled asynchronous worker queue with in-house auto-scaler
AWS Cloud Architecture
// 01 INGESTION TIER
Handheld Mobile Capture
iOS / Android / iPadOS (LiDAR)
5-10s video sweep & chunked multipart ingress
REST / JWT
Stateless REST API Tier
JAX-RS / Spring / MySQL RDS
Upload token validation & immediate job ticket issuance
Enqueues Payload Ticket
// 02 QUEUE BUFFER & AUTO-SCALING
Amazon SQS Message Queue
Decoupled Buffer + Dead-Letter Isolation
Buffers bursty daytime traffic; zero dropped scans
In-House Auto-Scaler
Queue Depth & Velocity Monitor
42% cost reduction vs AWS CPU lag; instant scale-down
Dispatches To Dynamic Pool
// 03 STORAGE & COMPUTE EXECUTION
S3 Multi-Tier Storage
Standard → Infrequent Access → Glacier
Cross-region replication; coordinate metadata kept live
Dynamic GPU Compute Pool
PyTorch / Open3D / CUDA Containers
Dense point cloud generation & sub-millimeter inference