Real-Time Damage Detection at Door

Visible damage at delivery is one of the clearest quality failures in last-mile logistics, and one of the hardest to prove later. The carton may be crushed, the seal broken, or the pallet wrap torn, but the only record is often a rushed photo, a vague exception code, or a customer complaint days later. By then, carrier liability windows have narrowed, chain-of-custody is fuzzy, and operations teams argue over whether the damage happened in transit, at the door, or after the driver left.

Real-time damage detection at the door closes that gap at the moment of handoff. A computer vision (CV) model reviews the delivery photos the driver already captures, classifies visible cargo defects, and writes structured, timestamped evidence into the delivery record. Each finding cites a defect class and the photo ID that supports it. If the cargo is not visible in the frame, the model returns empty rather than inventing a result. The driver still notes the exception in the normal workflow. The model does not replace human judgment; it makes the visual record usable for quality review and claims.

This sits downstream of warehouse checks such as damaged-goods vision check at pick, and alongside broader CV proof-of-delivery verification. Door detection is specifically about cargo condition at delivery, not signature capture or address confirmation alone.

The workflow starts when the driver completes proof-of-delivery capture. Images are sent to an on-device or edge-assisted CV pipeline that looks for cargo surfaces in the frame: cartons, crates, bags, totes, or pallet faces when those are visible. The model scores defect classes such as crush, puncture, tear, wetness indicators, open flap, broken seal, or skewed stack, depending on how the carrier trains and labels its taxonomy.

For each positive detection, the system writes a finding that includes:

• Defect class (controlled vocabulary, not free text)
• Confidence score and threshold outcome
• Photo ID of the supporting image
• Capture timestamp and delivery stop ID
• Bounding region or crop reference when available

If no cargo is visible, or confidence falls below the operating threshold, the output is empty. That empty result is intentional. It tells claim and quality teams that vision did not establish condition, so they must rely on the driver’s exception note and any customer follow-up. Hallucinated “no damage” labels from occluded or partial views create worse evidence problems than silence.

The driver path stays familiar. They still flag an exception when something looks wrong, refuse a delivery when policy requires it, or capture additional angles when the first shot is blocked by a porch wall, dusk lighting, or a customer who takes the package indoors before photos finish. Vision augments the exception record; it does not silently clear it.

Freight and parcel claims fail when the evidence pack is incomplete: missing photos, unlabeled damage type, or no proof that the image belonged to that stop. Door detection turns the photo set into a machine-readable evidence bundle the moment the stop closes.

A typical package includes the raw images, the model findings (defect class plus photo ID), stop metadata, and the driver’s exception note when present. That bundle feeds quality dashboards and, when severity warrants it, a freight claim filing agent that assembles the carrier claim with consistent attachments. Downstream exception routing can also consume the same signal via a delivery exception classifier, so “damaged at door” does not sit as an undifferentiated catch-all next to weather delay or wrong address.

Quality outcome measurement is straightforward. Track share of damage exceptions with at least one cited defect class and photo ID, time from delivery to claim-ready packet, and overturn rate when carriers challenge photo sufficiency. Empty-result rates are a process KPI, not a model failure by default: high empty rates usually mean capture SOP problems (too far, too dark, cargo already handed over) rather than CV alone.

Carriers and 3PLs rarely start from a blank model. Warehouse and yard operations already use industrial vision vendors such as Cognex and SICK for barcode, dimensioning, and defect inspection on conveyors and dock doors. Those systems prove that labeled defect taxonomies and photo-linked findings work under controlled lighting. Doorstep delivery is less controlled, but the evidence pattern is the same: detect, classify, cite the image, timestamp the event.

Last-mile platforms such as Onfleet already orchestrate stop workflows, photo capture, and exception codes. The practical integration is to attach CV findings to the existing POD object rather than inventing a parallel photo archive. Specialized inspections platforms such as Inspektlabs focus on vehicle and cargo damage assessment from images; logistics teams can reuse similar labeling discipline (defect class, photo ID, severity) even when the deployment surface is a handheld driver app instead of a fixed gantry camera.

The architectural choice that matters most is provenance. Every finding must resolve to a photo ID that still exists when a claim is audited six weeks later. Soft-deleted mobile uploads and overwritten “retake” images break that chain. Treat photo IDs as durable claim artifacts, with retention aligned to your longest carrier dispute window.

Visibility gate first. Run detection only when a cargo region is present. If the frame shows a door mat and a signature pad but no package face, return empty and prompt for a retake when policy allows. Do not auto-close as undamaged.

Driver exception remains authoritative for process. If the driver notes damage and vision is empty, escalate for human review. If vision flags crush and the driver notes nothing, surface a soft prompt before stop close. Dual signal (driver note plus CV finding) is the strongest quality record.

Taxonomy discipline. Prefer a short, auditable defect list over open-ended captions. “Crush / puncture / wet / open / other-visible” beats poetic free text that claims handlers cannot map to carrier codes.

Thresholds by lane and package type. Soft goods and rigid cartons do not share the same false-positive profile. Tune separately for high-value or claim-heavy lanes first, then expand.

Privacy and customer premises. Capture only what POD already permits. Crop to cargo when possible. Do not retain incidental faces or interior rooms beyond retention policy.

Feedback loop. When claims are upheld or denied, feed the outcome back into labeling and threshold review. Door detection improves when denied claims are classified as capture failure, taxonomy mismatch, or true model error.

Start with a shadow mode on a single region or high-claim SKU family. Log findings without changing driver UX. Measure agreement between CV defect class and later claim adjuster labels. Fix capture SOPs before you raise confidence thresholds.

Move to assistive mode: show the top defect suggestion and require the driver to confirm or override when cargo is visible. Keep empty results quiet unless a retake is cheap. Only then auto-attach findings to the delivery record for claim assembly.

Instrument the outcome as quality, not only claims cost. You want fewer ambiguous exceptions, faster evidence packs, and clearer separation between “damage visible at door,” “damage not visible,” and “customer allegation after delivery.” Real-time detection earns its place when every finding that ships into the claim file cites a defect class and a photo ID, and when silence means “not visible,” not “assumed fine.”