Home Expertise Case Studies Book Consultation

Case Studies

Deep technical proof of production deployments across industrial inspection and video analytics.

Case Study 00

Sentinel: Real-Time Traffic Monitoring Platform

Problem

Traffic authorities needed simultaneous monitoring of 100+ live RTSP camera streams with real-time event detection — including traffic signal violations, vehicle classification, occupancy, ANPR, and face recognition — without frame-drop or alert delay at production scale.

Approach

Built a distributed streaming pipeline using NVIDIA DeepStream as the primary GPU accelerated inference backbone, decoupled via Kafka for event routing and Redis for low-latency state management. Triton Inference Server handled model serving at FP16/INT8 precision for all concurrent models.

Architecture
100x RTSP Feeds
→
DeepStream Pipeline
→
Triton (TensorRT)
→
Kafka Event Bus
→
Redis State Store
→
Alert & Dashboard

Analytics modules running in parallel per stream:

Face Recognition — cross-view identity matching across multiple camera angles
ANPR — license plate detection, OCR, and vehicle owner lookup
Vehicle Classification — type, color, and make identification
Occupancy Detection — zone-level counting and threshold alerting
Traffic Signal Violation — red-light run detection with timestamp logging
Challenges & Solutions
Scaling to 100 simultaneous streams without GPU memory overflow → multi-process DeepStream bins with batching and stream mux optimization
Low-latency event delivery across modules → Kafka partitioning with per-stream topics and consumer groups
Model serving for 5+ concurrent analytics tasks per stream → Triton dynamic batching with per-model concurrency controls
Real-time face matching across non-overlapping views → FAISS-backed feature vector store for sub-millisecond similarity search
Outcome
100
Live RTSP Streams
30 FPS
Maintained per Stream
5+
Parallel Analytics Modules
<500ms
Alert Latency
Case Study 01

High-Speed Bottle Inspection

Problem

Manual inspection bottleneck in high-speed manufacturing — could not keep pace with 200 units/minute production rate. High false-acceptance rate causing quality escapes.

Approach

Real-time anomaly detection pipeline with GPU-accelerated inference. Unsupervised + supervised hybrid approach to handle label-scarce manufacturing environment.

Architecture
Camera Feed
→
Preprocessing
→
Anomaly Detection (TensorRT)
→
Defect Classification
→
Accept / Reject
Challenges & Solutions
Maintaining inference latency under 200ms at 200 units/min
Low defect sample availability → used unsupervised + semi-supervised methods
Real-world lighting and surface variation
Outcome
>95%
Inspection Accuracy
<1.5%
False Acceptance Rate
180ms
Inference Latency
~90%
Manual Effort Reduced
Case Study 02

Multi-Camera Drive-Thru ReID (Wobot.ai)

Problem

Tracking individual vehicles across non-overlapping camera views in drive-thru lanes to measure dwell time and wait-time patterns. Identity switches caused analytics to be unreliable.

Approach

Custom multi-camera ReID pipeline with tracker integration and iterative retraining on real-world failure cases.

Architecture
RTSP Feeds (1..N)
→
Detection
→
Single-cam Tracker
→
ReID Matching
→
Cross-cam Graph
→
Analytics Engine
Challenges & Solutions
Occlusion, lighting variation, vehicle appearance similarity
Cross-camera view gap (non-overlapping FOV)
Scale: 500+ live cameras across 20+ locations
Outcome
2–5%
Identity Mismatch (Down from 15%)
500+
Live Cameras
20+
Locations in Production
Case Study 03

Wrist-Watch Defect Detection

Problem

Manual 360° visual inspection of wrist-watch cases taking 30 seconds per unit — creating a production bottleneck and returning inconsistent results due to inspector fatigue and subjective judgment. Multiple surface types (bezel, case back, lugs, crystal) each required independent scrutiny, making automation non-trivial.

Approach

Dual-camera synchronized rig engineered to capture all watch surfaces in a single pass. A defect detection model was trained to recognize micro-scratches, dents, coating failures, and misalignments per surface zone. Hardware trigger synchronization eliminated frame drift between the two cameras.

Architecture
Dual Camera Rig
→
HW-Sync Capture
→
Surface Alignment
→
Defect Detection Model
→
Per-surface Defect Map
→
Pass / Fail Decision
Defect Types Covered
Micro-scratches on bezel and case back
Coating delamination and discoloration
Dents, deformations, and edge chips
Crystal cloudiness and surface contamination
Lug misalignment and finishing inconsistencies
Challenges & Solutions
Dual-camera synchronization without frame drift → hardware trigger with sub-millisecond sync tolerance
Multiple surface types with different defect signatures → per-surface specialized detection heads
Cycle-time constraint (must be faster than manual 30s) → parallel inference on both camera feeds with batched model execution
High-gloss reflective surface making consistent imaging difficult → custom diffuse lighting rig engineered to eliminate specular reflections
Outcome
8s
Per Unit (was 30s)
73%
Cycle-time Reduction
360°
Surface Coverage
5+
Defect Types Detected