AI-Based Camera Development with NXP i.MX, Firmware & PCB Design

Client: Industrial automation startup (North America)
Project Goals: Create a custom smart camera to detect defects in fast-moving production lines (at 120 units/minute). Results:Deployed a production-ready system to 3 factories, achieving 99.7% accuracy in detecting defects.

The production line already used human inspectors to check for quality and detect defects. The inspectors sometimes missed 15% of defects (wrongly placed labels, surface scratches, and missing parts) due to their inability to quickly look at the images as they moved through the line.

Both commercial machine vision options did not meet the requirements of the client due to one of the following reasons:

• Too expensive
• Too large (could not fit in the 100mm x 80mm available space)
• Missing specific features (need for custom lighting control, edge AI, and real-time cloud alerts)

• The common area lights do not turn off and waste energy
• An overflow of the water tank went unnoticed until it started flooding into the common room
• Misunderstanding with the monthly bills for utilities between members because there is no meter for each unit separately.
• There is no early warning system for gas leaks or electrical issues.

The customer required an AI-enabled compact and custom-built smart vision camera that must:

• Collect high-resolution photographs (5MP) at a rate of 120 images per minute;
• Perform real-time defect analysis at the edge without requiring the cloud for its primary functions;
• Provide control over its integrated LED lighting (utilising strobing for production line synchronisation);
• Send alerts and images of defects to a cloud-based server for analytical purposes;
• Continuously operate 24/7 in a rugged industrial environment (0oC - 50oC, dusty).
• Connect to a factory Programmable Logic Controller (PLC) for use as part of a rejection mechanism.

Adequate Infosoft was selected to develop the entire system through consulting, engineering, and manufacturing capabilities through hardware design and firmware development and producing the software for the cloud platform along with training the artificial intelligence models, this was all to be completed in a timeframe of less than 4 months.

Hardware & PCB Design

We designed a custom, compact PCB integrating all required components into a 90mm x 70mm form factor.

Hardware Architecture

Key Hardware Specifications

Component	Selection	Rationale
Image Sensor	OnSemi AR0521 (5MP, 2592x1944)	Global shutter (critical for moving objects), excellent low-light performance
Main Processor	NXP i.MX RT1060 (Cortex-M7 @ 600MHz)	Real-time performance, industrial temperature range, rich peripherals
AI Accelerator	Google Coral Edge TPU (4 TOPS)	4 trillion operations per second for real-time inference
Memory	512MB DDR3 + 8GB eMMC	Frame buffering + firmware storage
Connectivity	Gigabit Ethernet (W5500) + optional 4G	Factory network integration + remote access
I/O	4x optically isolated inputs, 4x relay outputs	PLC trigger input + reject output
Illumination	High-power LED driver (up to 2A, PWM controlled)	Synchronized strobe lighting for clear images
Power	24V DC industrial input, 5V/3.3V onboard regulation	Standard factory voltage
PCB Layers	8-layer, controlled impedance, ENIG finish	Signal integrity (MIPI, DDR), industrial reliability

Critical PCB Design Challenges & Solutions

Challenge	Our Solution
MIPI CSI-2 signal integrity	Matched trace lengths (0.5mm tolerance), 100Ω differential impedance, ground shielding between pairs
High-speed DDR3 routing	Fly-by topology with on-board termination, length matching within 2mm
Thermal management (Coral TPU)	4x4mm thermal pad under device, 9 thermal vias to bottom-side copper pour, aluminum enclosure contact
EMI/EMC compliance	Split digital/analog grounds, ferrite beads on power rails, shielded enclosure with gasket
Industrial environment	Conformal coating on all PCBs, wide-temp components (-40°C to 85°C rated)

PCB Layout Statistics:

• 8 layers, 90mm x 70mm
• 1,247 components, 2,845 nets
• 3 months of layout effort (including 2 re-spins for optimization)
• Passed thermal, signal integrity, and EMI simulations before fabrication

Firmware & Embedded Software

Our firmware team wrote efficient, real-time code running on the i.MX RT1060 and Coral Edge TPU.

Firmware Architecture

Key Firmware Features Implemented

Feature	Implementation Details
Trigger-to-image latency	<50ms from PLC trigger to captured image (optimized sensor register config + double-buffering)
AI inference time	<30ms per frame (YOLOv8-tiny on Coral TPU)
Real-time detection	120 units/minute = 500ms per unit → 80ms spare margin
Image capture	5MP at 15fps, auto-exposure with strobe synchronization
Edge storage	8GB eMMC circular buffer (last 10,000 defect images)
Network protocols	MQTT (telemetry), FTP/HTTP (image upload), Modbus TCP (PLC integration)
OTA updates	Firmware, Edge TPU model, and bootloader updates over Ethernet
Watchdog	Hardware watchdog timer with 5-second timeout, auto-reboot on hang
Self-diagnostics	Temperature monitoring, frame drop detection, network health checks

Codebase Statistics:

• 35,000+ lines of C/C++ firmware
• 5,000+ lines of Python (AI model training + cloud integration)
• 15 FreeRTOS tasks (prioritized: camera capture = highest, network = lowest)

AI Model Development & Training

We trained a custom defect detection model for the client's specific product line.

Model Pipeline

Defect Classes Detected

Class	Description	Training Samples
label_misaligned	Label rotated or shifted >2mm from position	4,200
surface_scratch	Any scratch >1mm visible under strobe light	3,080
missing_component	Required part absent from assembly	3,500
contamination	Dust, debris, or foreign material	3,500

Model Results:

• 0.5 mAP (mean Average Precision): 97.2
• Inference time on Coral TPU: 28 ms (5 MP image)
• False Positive Rate: 0.8% (after twice retraining to lower from 3.2%)
• False Negative Rate: 0.3% (missed defects)

Training Setup:

• Google Colab Pro + Tesla T4 GPU
• 5000 Training epochs (3,200 early stop)
• Augmentations: rotate +/- 5°, brightness +/- 20%, and add noise.

Cloud Platform & Backend

We built a complete cloud infrastructure for fleet management, data analytics, and remote monitoring.

Cloud Architecture (AWS)

Cloud Features Built

Feature	Implementation
Device provisioning	Zero-touch onboarding (certificate-based, QR code scan)
Real-time telemetry	MQTT keep-alive, last will (offline detection), 5-second latency
Defect image storage	S3 with lifecycle (30 days hot, 90 days cold, delete after 1 year)
Alert system	SNS → SMS/Email (production manager, maintenance team)
Analytics dashboard	Defect rate trends, top defect types, machine OEE (Overall Equipment Effectiveness)
Fleet OTA updates	Gradual rollout (10% → 50% → 100%) with automatic rollback
Audit logging	All configuration changes logged to DynamoDB
Secure access	IAM roles, API keys, JWT authentication

Cloud Usage Metrics (3 plants, 25 cameras, 6 months):

• 4.5 million images processed
• 2.3 TB storage consumed
• 99.95% uptime (AWS multi-AZ deployment)
• Average API response: 120ms

Factory Integration & PLC Communication

The smart camera needed to talk to existing factory equipment.

Integration Architecture

PLC Protocol Implemented:

• Port 502 Modbus TCP
• Trigger input: camera listens for specific holding register (write a ‘1’ to capture)
• Result output: camera writes defect code to holding register (pass = 0; label = 1; scratch = 2; etc.)
• Handshake: Camera already written 255 to clear register once rejection mechanism is confirmed as complete

Latency measured:

• PLC Trigger to camera capture = 15ms (Ethernet Switching Time)
• AI inference to reject signal = 30ms
• Reject mechanism activation = 50ms
• Total line interruption per defect = <100ms (well below line speed tolerance)

Prototyping Phases

Phase	Activity	Duration	Outcome
Phase 1	Breadboard proof-of-concept (i.MX RT eval board + USB camera)	3 weeks	Verified 120 units/min feasibility
Phase 2	Custom PCB prototype (2 units, hand-assembled)	6 weeks	Board bring-up successful, MIPI working
Phase 3	Pre-production run (25 units, assembly house)	8 weeks	All units passed factory acceptance test
Phase 4	Pilot deployment (2 cameras, 1 production line)	4 weeks	99.7% accuracy in real conditions
Phase 5	Full deployment (25 cameras, 3 plants)	6 weeks	Live production, 24/7 operation

Testing Results

Test Category	Metric	Result
Accuracy	Defect detection rate	99.7% (vs human 85%)
Latency	Trigger to output	48ms average (well under 500ms requirement)
Throughput	Units processed per minute	122 (exceeds 120 target)
Uptime	Operational availability	99.5% (planned maintenance excluded)
Temperature	Internal PCB temp at 40°C ambient	62°C (within spec)
Vibration	IEC 60068-2-6 (10-500Hz)	Passed no solder joint failures
EMC/EMI	EN 61000-6-2 (industrial immunity)	Passed all criteria
Power consumption	24V @ 350mA average, 850mA peak (LED strobe)	8.4W avg, 20W peak

Category	Deliverable	Format
Hardware	Custom PCB design files	Altium schematics, Gerber, BOM, pick-and-place
Hardware	10 prototype units + 25 production units	Assembled, tested PCBs
Hardware	Enclosure mechanical design	STEP files, 2D drawings
Firmware	Production firmware binary	.bin + OTA update package
Firmware	Source code (escrow)	C/C++ (FreeRTOS, drivers, application)
AI Model	Trained YOLOv8 model + Edge TPU version	.pt, .tflite, .tflite compiled for TPU
AI Model	Training pipeline scripts	Python scripts + Jupyter notebooks
Cloud	AWS CloudFormation templates	Infrastructure as code (repeatable)
Cloud	Backend API + Lambda functions	Python (serverless)
Cloud	Web dashboard (React.js)	Deployed on AWS S3 + CloudFront
Integration	PLC communication library	Modbus TCP example code
Documentation	Hardware design guide	45-page PDF (schematics, layout, test points)
Documentation	Firmware API reference	60-page PDF
Documentation	Integration guide (PLC, network)	25-page PDF
Documentation	User manual (operators)	20-page PDF
Documentation	Maintenance guide	15-page PDF
Testing	Validation report	80-page PDF (all test results)
Training	Operator training session	4-hour on-site + video recording
Training	Technician training session	8-hour on-site + hands-on

Quantitative Results (After 6 Months of Production)

Metric	Before	After	Improvement
Defect detection rate	85% (human)	99.7%	+14.7%
False rejects (good units rejected)	3%	0.5%	-83%
Annual customer returns	1,200 units	140 units	-88%
Inspection labor cost	$180,000/year	$30,000/year	-83%
Line downtime (inspection-related)	15 hours/month	2 hours/month	-87%
ROI epriod	—	9 months	—

Qualitative Outcomes

Customer Trust

A significant retailer decreased audit intervals when the defect percentage decreased.

Data driven decision making

Daily defect reports established root causes (specific shifts, machine settings).

Scalability

The same camera platform has been implemented for two additional product lines.

Predictive Maintenance

Defect trend analysis allows for forecasting when a packaging machine will require calibration.

What Worked Out

• Co-design of Hardware and Software Reduced Integration Time by 2 Months
• Coral TPU Was an Appropriate Selection, Inference Speed Exceeded Our Need
• Image Captured using Double-Buffering Stayed at 120 Units/Minute Without Frame Drops
• Modbus TCP Was Less Complicated Than Originally Anticipated and Did Not Require an Additional Custom PLC Driver

Challenges Overcome

Challenge	Solution
MIPI signal integrity on first PCB revision	Re-routed with matched lengths, added ground shielding, second spin passed
Coral TPU thermal throttling at 50°C ambient	Added thermal pad to enclosure + forced air (small fan)
False positives on shiny surfaces	Added polarized lens + adjusted lighting angle
OTA update corruption risk	Dual-bank flash + signature verification + rollback timer

Adequate Infosoft has provided a full end-to-end, custom smart camera solution that replaced the client's quality control process.

We created all of the hardware, developed the firmware, trained the AI model, developed the cloud platform, integrated the solution with factory PLCs, and provided deployment support on production lines.

Some of the most significant accomplishments of this engagement include:

• Defective product detection at 99.7% accuracy
• More than 120 units per minute of throughput
• 9-month payback on investment
• Deployment at 3 manufacturing facilities
• No hardware failures for 6 months
• Expansion into two additional product lines from customer

Get Free Technical Consultation On Your Custom Smart Camera Contact Adequate Infosoft Now We provide all kinds of Smart Camera solutions including AI Cameras, Firmware Development, PCB Design, Embedded Systems, Prototype Development and Final Production Development for your custom Smart Camera.