Pillar 01 · AI Measurement
Two measurement systems, not one dependency
Camera AI proves the exposure happened. Mobile and geofence analytics prove who was exposed and what they did next. Each layer covers the other's blind spot, and every headline number is reported with a cross-check between them.
Layer A
Camera AI at the asset
Edge vision on the structure itself. It answers the question a media buyer actually asks first: did real traffic pass this board, in the right direction, with the board in a viewable state?
- Vehicle count — Continuous per-lane counting, 24/7, at the board itself.
- Class recognition — Motorcycle, car, SUV/pickup, jeepney/bus, truck — each with its own occupancy factor.
- Make / model recognition — Vehicle mix as a proxy for income and lifestyle cohorts — no plates, no owners.
- Traffic direction — Northbound vs southbound so only facing traffic counts.
- Estimated visibility — Obstruction, lighting and print condition scored into a visibility index.
- Dwell / exposure time — How long the board stays in view under real congestion, not a rate-card assumption.
Layer B
Mobile & geofence analytics
Consented panel data, geofenced around the viewshed and around advertiser locations. It answers the second question: how many distinct people, how often, and did it move anything?
- Estimated unique audience — De-duplicated reach behind the raw impressions.
- Repeat exposure & frequency — How many times the same cohort passes the site.
- Audience movement — Trip purpose mix: work, home-bound, retail, leisure.
- Store visitation — Modelled lift in visits to advertiser locations post-exposure.
- Website / app conversions — Exposed-cohort sessions and completed actions.
- Campaign attribution — Exposed vs control uplift, attributed value and ROAS.
Why both
Neither system is credible alone
Exposure truth
Camera counts are a census at the asset. They anchor impressions to something physically observed instead of a modelled traffic estimate.
Audience truth
A camera cannot tell you the same rider passed eleven times. Panel-based geofence data supplies de-duplication, frequency and downstream behaviour.
Cross-validation
Panel-implied exposure (unique audience × frequency) is compared against camera-counted opportunity-to-see. A small delta raises confidence; a large one flags a sensor or panel problem.
Compliance
Built to the Data Privacy Act of 2012
Measurement is environment-level by design. The system is architected so that personal data is never the unit of analysis — which keeps processing proportionate, minimises retention, and makes the whole stack defensible to the National Privacy Commission and to an advertiser's own compliance team.
No individual profiling
Camera AI stores counts, classes and directions — not faces, not plates, not identities. Frames are processed at the edge and discarded; only aggregate counters are retained.
Cohort aggregation only
Geofence outputs are panel-based cohort estimates extrapolated to corridor level. No single device is ever reported, exported or resold.
k-anonymity threshold
Any cohort cell below the minimum panel threshold is suppressed rather than reported, so no result can be narrowed down to one person.
Consent basis
Panel data is sourced only from providers with documented opt-in consent for location analytics, with contractual warranties covering downstream use.
Retention limits
Raw sensor frames: not retained. Aggregate counters: retained for the campaign term plus audit period. Panel extracts: deleted on schedule after reporting.
Audit trail
Every report generation, export and configuration change is written to an immutable activity log for review by the Data Protection Officer.
The re-identification boundary
Anonymous exposure data is never resolved back to a person. Linking a passer-by to a mobile advertising ID, an app registration, a purchase or a messaging account would convert environment-level measurement into personal-data processing without a lawful basis, and is prohibited by design — not merely by policy.
Permitted
- Corridor-level counts, classes, direction and dwell from camera AI — no faces, no plates.
- Cohort-level reach, frequency and movement from a consented panel, above the k-anonymity floor.
- Modelled store-visit and conversion lift reported as aggregate percentages.
- Clean-room matching against the advertiser's own consented first-party data, aggregates out only.
Prohibited
- Resolving an anonymous passer-by to a mobile advertising ID, SIM, or device fingerprint.
- Joining exposure data to app sign-ups, purchase records, loyalty accounts or messaging identities.
- Retargeting, scoring or building profiles of individuals seen by a camera.
- Selling, exporting or enriching device-level location data in any form.
- Reporting a cohort cell small enough to single out one person.
Where an advertiser wants outcome matching, it happens in a clean room on the advertiser's own consented first-party data, with a k-anonymity floor applied before any result leaves the enclave. Only aggregates come back — never a record, never an identifier.
Clean-room matching
How outcome matching happens without re-identification
ADVERTISER SIDE CLEAN ROOM (enclave) MEDIA OWNER SIDE
--------------- -------------------- ----------------
First-party CRM Camera AI counters
(consented) (no faces, no plates)
| |
| SHA-256 in browser | aggregate
v v
hashed keys -------------> [ policy engine ] <------------- exposure cohort
CR-01 identifier class (consented panel,
CR-02 declared purpose corridor level)
CR-03 aggregate egress only
CR-04 no camera-detection join
CR-05 consent attestation
CR-06 k-anonymity floor k >= 50
|
allowed | blocked -> no output, audited
v
[ cohort match ]
|
v
aggregates only: matched count,
exposed vs control, lift %
cells with n < 50 suppressed
|
v
append-only audit entry
(advertiser, decision, rules applied, timestamp)Hashes are compared inside the enclave and discarded. Nothing crosses back except counts and rates, and any cell containing fewer than 50 people is suppressed rather than published.
Roles. The media owner acts as personal information controller for its own operational data; panel providers remain controllers of their consented device data and supply only aggregate outputs, making them the source of a lawful basis rather than a liability to inherit. Data sharing agreements, a privacy impact assessment per deployed corridor, and a named Data Protection Officer complete the compliance posture.