AI Detection of Address for Fleetoptics

Key results

• 77.4% — Address verification accuracy
• 15,000+ — Images validated in production runs
• 2 months — Kickoff to production delivery
• 4 engines — YOLO + PaddleOCR + Google OCR + Tesseract

TL;DR

• Drivers upload parcel photos directly from delivery sites, and AI extracts address details automatically without delaying drivers.
• A custom ensemble was built because any single OCR failed on portions of real-world delivery imagery.
• The stack combines YOLO, PaddleOCR, Google OCR, and Tesseract to survive blurry, low-light, compressed, and under-processed Android/iPhone captures.
• With 77.4% address verification accuracy on 15,000+ images, the flow reduced dependence on manual Virtual Assistant review while staying practical for daily dispatch operations.

Project Overview

Fleetoptics needed a reliable, automated way to verify delivery addresses from phone photos taken in unpredictable field conditions and prevent mis-deliveries at scale.

Drivers took parcel photos with Android and iPhone devices in poor lighting, extreme angles, and motion blur. Historically, teams relied heavily on manual review and Virtual Assistant follow-up to validate addresses.

We built a custom OCR ensemble pipeline because no single engine could cover all edge cases. The system captures address regions, runs multiple OCR passes, normalizes noisy output, and validates through internal records before surfacing a delivery decision.

• Automated extraction of address text from real-world delivery imagery, including phone-captured edge cases
• Custom ensemble voting across models to recover text when one engine misses or corrupts output
• Location matching engine against company address database
• Fast status generation for driver and operations workflows

Business Challenge / Problem Statement

Manual address checks were slow, inconsistent, and resource-heavy.

VA teams spent significant time reviewing photos and confirming parcel location details because single OCR output alone was too fragile.

Android/iPhone images were heavily compressed, under-processed, blurry, overexposed, and often clipped, causing severe OCR drift in a traditional pipeline.

Varying labels, hand-written additions, font styles, and sticker placements required a defense-in-depth approach to extraction.

Delivery confirmation needed to be faster to keep drivers moving efficiently.

• High Manual Load — Large QA and support volume reduced driver and coordinator efficiency.
• Address Extraction Drift — Single OCR methods underperformed on inconsistent imagery, low-res captures, and mixed language/label styles from driver phone photos.
• Verification Delay — Confirmation lag created slower final delivery closure and operational overhead.

Objectives & Goals

The system had to be accurate, fast, and practical in real field conditions.

Create an automated, end-to-end address verification flow from photo upload to database confirmation.

Build a custom multi-engine ensemble so each OCR can cover the others' blind spots instead of failing as a single point of failure.

Optimize the full pipeline for production-grade noise: underexposed, motion-blurred, compressed, and partially cropped delivery photos from Android and iPhone devices.

• Robust Address Reading — Extract house/street numbers consistently from real delivery photos, including partially visible signs, handwritten annotations, and inconsistent print styles.
• Model Ensemble Accuracy — Use PaddleOCR, Google OCR, Tesseract, and YOLO together as a staged ensemble to increase recall, precision, and stability on noisy delivery images.
• Mobile-Photo Resilience — Handle real-world capture variability across device quality, angle drift, motion blur, low light, and compression so output quality degrades gracefully.
• Operational Matching — Validate recognized addresses against Fleetoptics records and return clear pass/fail outcomes.

Key Features & Capabilities

A practical AI workflow for delivery operations teams.

Photo Intake and Localization

Driver images are ingested, then YOLO localizes likely text regions and filters background noise so each OCR pass focuses on the most relevant parts of the parcel photo.

Address Extraction Stack

• PaddleOCR: Primary detector for dense, structured address text when labels are crisp enough for high-confidence reads.
• Google OCR + Tesseract: Cross-engine fallback for underexposed, noisy, and compressed captures where one model alone fails.

Address Normalization

• House and street parsing: Structured extraction of house and street components with explicit handling for partial reads, missing digits, and OCR segmentation noise.
• Cross-engine adjudication: Confidence-aware voting resolves disagreements between OCR engines to reduce one-model false positives.
• Standardization: Cleansed formatting to reduce false mismatches from noisy text.

Database Validation

Parsed address data is checked against company records through API to confirm correct delivery location.

Results & Impact

The platform automated a large portion of delivery address checks that were previously manual and delayed final delivery closure.

Drivers moved through delivery workflows faster with immediate confidence checks available from AI validation, even on poor Android and iPhone captures.

The custom ensemble model architecture reduced false misses and false matches by compensating for model-specific weaknesses across thousands of noisy field images.

What our clients say

Address verification became much faster and far more dependable once the custom OCR ensemble was deployed across low-quality delivery photos.

Fleetoptics Operations Team — Logistics Operations

Technologies & Tools Used

Vision and OCR stack for logistics automation

An ensemble of YOLO for localization plus multiple OCR engines for resilient extraction when phone photos are blurry, underexposed, compressed, or partially framed.

API-based validation integrates directly with internal address records for near real-time verification.

Technology stack

• PaddleOCR — Primary OCR model for robust text extraction in field images.
• Google OCR — Secondary OCR model to improve recall and confidence.
• Tesseract — Fallback OCR for variable fonts and formatting styles.
• YOLO — Object detection for locating relevant text regions in parcel photos.
• Fleetoptics API — Internal address database matching and delivery-record verification.