Fleet AI and bus occupancy: turn connected devices into reliable decisions
A detailed review of tracking, passenger counting, cameras and maintenance data, with practical checks and human decision points.
Start with a decision
More devices do not automatically create better knowledge. Define the decision: explain delays, improve capacity allocation or close maintenance actions. Vehicle location answers a different question from passenger count. Connect each signal to an action and an accountable person.
A delay report should lead to a timetable or boarding review; a maintenance alert should lead to an inspection record. This is an operating model whose equipment and integrations depend on the project, not a claim that every vehicle has identical technology.
What occupancy research shows
Dib and colleagues combined fare-collection and passenger-counting data where sensor coverage was incomplete, evaluating their method on French networks. The important point for operators is that sources can be biased or missing and estimates require assessment. Overseas model output is not a ready-made measurement of a Saudi fleet.
Our practical interpretation distinguishes direct measurements, calculated estimates and missing values. Displaying that distinction helps avoid false confidence. Review early capacity estimates against observed counts from a relevant sample of journeys.
Create one journey record
Use a stable journey identifier linked to vehicle, shift, route and actual times. Align vehicle identifiers across tracking and maintenance. For API integrations, record refresh time and errors; for imported files, record the reporting period, origin and upload date.
Direct integration is not always available. Structured files and validated manual entries can provide a useful starting point. Prevent duplicate journeys, review missing fields and expand connections after checking accuracy rather than replacing every system at once.
Validate passenger counting
Boarding patterns, crowding and device angles can affect counts. Compare sensors with manual observations covering different doors and peak conditions. A supplier accuracy claim is not a validation of your own operating environment.
For a hypothetical example, 36 passengers on a 49-seat bus represent about 73.5% seat occupancy at that point. This is neither a daily average nor financial utilisation. Define measurement points, passenger changes and periods before comparing routes.
Camera alerts need review
Driver-monitoring cameras may support fatigue or distraction alerts, depending on equipment and conditions. They work alongside training, suitable shifts and safe stopping arrangements. An alert is not a medical diagnosis or a replacement for operating responsibility.
Agree access, purpose, retention and review of false alerts before deployment. Minimise personal data and avoid showing passenger identities in general dashboards. Applicable privacy requirements and approvals must be assessed for each project.
- 1Field devicesLocation, counts, alerts
- 2Unified recordJourney, vehicle, time
- 3ValidationMeasured or estimated?
- 4Action & reviewOwner, action, outcome
- On-time performance
- In-window runs ÷ included runs
- Seat occupancy
- Passengers ÷ seats at a defined point
- User cost
- Defined cost ÷ actual users
- Vehicle availability
- Available time ÷ required time
Prediction needs a sound maintenance record
An abnormal reading can justify inspection without proving that a component will fail at a precise time. Record defects, repairs, inspections, mileage and timestamps. Without that context, repeated alerts may add work without improving continuity.
Start with understandable rules for thresholds, recurring faults and overdue checks. Evaluate predictive models only when suitable evidence exists. Scheduled preventive maintenance remains important, and an automated score must not justify operating an unsafe vehicle.
Define four useful measures
On-time performance needs an agreed arrival window and a documented denominator. Occupancy requires a defined point or period. Passenger unit cost depends on included expenditure and actual users. Availability requires recorded service and downtime rather than a vehicle inventory alone.
The accompanying diagram explains these definitions without inventing company performance percentages. Clear definitions help customers understand changes and identify the action each measure supports. Illustrative dashboards should be labelled as demonstrations.
Run a focused pilot
Choose one question, a limited group of vehicles and a representative period. Record the baseline, then assess accuracy, usability and response, alongside device, connectivity and integration costs. Define continuation criteria before the trial.
The output may recommend additional equipment, a revised workflow or better source data. A useful pilot explains benefits, limits and scaling requirements so technology becomes part of service delivery rather than a collection of impressive screens.