To measure passenger traffic, a train station does not need a new set of sensors. The cameras already installed for security purposes are sufficient in most cases, provided the right views are selected and each is assigned a specific metric. The process therefore involves less adding equipment and more defining, camera by camera, the area being monitored and the expected metric.
Three categories of metrics cover most operational needs: passenger density in an area, the flow rate through a counting line, and the travel time between two points. A station that has these three metrics for about ten well-chosen zones can detect congestion before it blocks a staircase or platform access. XXII’s CORE platform is based on this principle: it analyzes the video feeds from existing cameras and extracts these indicators, without facial recognition.
The list is short. Main entrances and swing doors provide data on incoming and outgoing passenger flow, and thus the building’s overall occupancy at any given moment. Staircases, escalators, and elevators are the points where high density becomes dangerous, because the flow is confined there and turning back is impossible. Waiting areas in front of departure screens concentrate passengers who do not yet know which platform to head toward, and serve as the best early indicator of a peak in platform access.
Next come the lines: ticket windows, self-service kiosks, and access control points where they exist. Then there are the transfer corridors, where travel times reveal a slowdown well before density becomes visible. A medium-sized station meets these needs with ten to twenty zones. Increasing the number of zones beyond that complicates operations without improving decision-making.
A camera is suitable when it views the zone from a sufficient height and at a downward angle. A view that is too low causes people to overlap and makes it impossible to distinguish them, which skews the count as soon as density increases. It is precisely during the times of greatest interest to the operator that measurement accuracy deteriorates the most.
Other points to check include the actual resolution available on the secondary feed, the frames per second rate, the stability of the frame, and nighttime lighting. A motorized camera controlled by security personnel is not suitable for counting: every movement invalidates the counting zone. It is still possible to keep it for security purposes and add a fixed camera dedicated to counting at the same location.
In a heterogeneous system, a preliminary audit saves time. Classifying cameras into three groups—usable as-is, usable after adjustment, and unusable—prevents the discovery at the end of the project that the most critical views are unusable.
On-site processing, directly near the cameras, prevents images from leaving the station’s network. Only the metrics—that is, numerical data—are sent to the monitoring system. This architecture reduces the required bandwidth, simplifies communication with IT security teams, and minimizes the risk associated with personal data.
Centralized processing remains a viable option when the network allows it and when the operator wishes to pool computing power across multiple sites. The decision is rarely based on technical factors alone: it depends on internal policies regarding the transmission of images and the ability of local teams to maintain the equipment.
A measure that triggers no action serves no purpose. Integration occurs on two levels. In real time, density thresholds feed into the existing monitoring system—whether a command center or hypervision platform—and trigger an alert when an area exceeds a predefined level. The operator receives the information, including the area’s name and the trend over the past few minutes, rather than an image that requires interpretation.
Historical data is used to adjust transportation schedules and staffing levels. Comparing the hourly profile of a typical Tuesday with that of a peak departure day allows us to deploy reception agents where they can actually make a difference. These two uses require different data exports: real-time data goes through a machine interface, while analysis goes through a data warehouse.
The only reliable method is to establish a baseline based on actual field data. An agent manually counts passengers using a recorded video sequence, covering several representative time slots—including at least one during peak hours and one at night. The discrepancy between the manual count and the automatic measurement provides the actual confidence level for that specific camera.
This verification must be repeated after every change to the facility: a new camera, a lens change, or a redesign of the area. A count calibrated for a concourse before construction work is no longer valid after a ticket counter is moved. Without this periodic check, the measurement drifts without anyone noticing.
A train station is a public space. Video surveillance systems there are governed by the Internal Security Code, which requires prefectural authorization and sets a maximum retention period of one month for footage. The addition of analytical processing to these video streams does not exempt the system from this framework; rather, it is an additional requirement.
Regarding personal data, the GDPR applies as soon as a person is identifiable in the image. Measuring density and throughput does not involve identifying anyone, which makes it possible to design the processing so that it produces only aggregated data. Facial recognition constitutes biometric data, the processing of which is prohibited in principle unless there is a legal exception: it has no place in a traffic-counting project.
A data protection impact assessment is required whenever the system systematically monitors a large-scale, publicly accessible area. Finally, if the system also covers workspaces, the consultation with employee representatives required by the Labor Code must be conducted before the system is put into service.
A reasonable rollout begins on a limited scale: three to five areas, selected from those where operational decisions are already being made manually today. The goal of the first stage is not to cover the entire station, but to verify that the system aligns with what staff members observe and that it is deployed in the right place at the right time.
The question to ask before any expansion is simple: What decision has changed as a result of this metric over the past month? If there is no answer, the problem lies in how the information is routed or in the choice of zones—not in the technology. Expanding the system before answering this question is tantamount to multiplying indicators that no one looks at.