Intelligent video surveillance: from signal to action

See, interpret and alert: three different functions

The camera captures, the algorithm tests for a condition and the operator interprets. A box drawn around a silhouette is not proof of intrusion; a model classified a shape at a given threshold. Time, credentials, trajectory, sensors and adjacent views supply missing context.

• Observe: provide usable imagery in the site's real operating conditions.
• Detect: flag a configured rule or anomaly.
• Qualify: compare the alert with several pieces of evidence.
• Decide: apply an instruction approved by an authorised person.
• Document: retain only what is necessary and auditable.

One security layer, never an absolute guarantee

A meta-analysis covering forty years of evaluations associates CCTV with a modest but significant reduction in crime, more pronounced in car parks and residential areas. Active monitoring and complementary measures improve outcomes. No one can promise that a camera will prevent a break-in; a targeted, integrated and evaluated system is a defensible proposition. This case is also reported in CCTV Surveillance for Crime Prevention: 40-year meta-analysis (Office of Justice Programs).

Visible-light cameras: detail, colour and understanding

A visible camera contributes colours and detail: an open door, clothing, direction of travel, a carried object and perhaps a number plate if framing permits. Lighting, contrast and shutter speed matter as much as resolution. A wide angle covers more ground but leaves fewer pixels on each target.

• Fixed dome for a stable, discreet view.
• Bullet camera to make direction and perimeter apparent.
• Multi-sensor camera for several angles with fewer blind spots.
• Motorised PTZ for investigating events, without replacing fixed views.
• Near infrared at night, with attention to reflections and insects.

Thermal cameras: detect contrast, not a face

Thermal imaging represents infrared radiation from surfaces and may reveal a person or vehicle in darkness. Rain, fog, warmed surfaces, reflections, vegetation and apparent target size still influence the result. It may show that a person-like shape is moving without revealing who they are. Where presence alone matters, that limitation can protect privacy. A visible camera or second sensor can then provide the detail required for verification. The stated specifications are detailed in Q1972-E thermal camera and edge analytics (Axis Communications).

Visible and thermal together, supported by complementary sensors

Along a dark fence, thermal imaging triggers while optical imaging qualifies. At a gate, access control explains a presence; on a loading bay, a sensor confirms the trajectory. The point is to cross-check independent signals, not accumulate devices.

Detect an event without profiling a person

Protecting a perimeter rarely requires knowing who walks beside a fence. Temporary local classification can simply open a clip. Swiss Federal Railways illustrates minimisation in another context: its 3D footfall sensors transmit density, speed and dwell-time measurements without images or individual tracking. This is not security surveillance, but it is purpose-led design.

Biometric identification creates another level of risk

Comparing a face with a database is not an innocuous extension. The FADP classifies biometric data that uniquely identifies a person as sensitive personal data. Necessity, proportionality, notices, errors, security and alternatives require careful assessment. A badge, zone rule and human verification are often sufficient.

Why cameras make mistakes

Vibration, rain, spiders, shadows, reflections, animals, tarpaulins, obstruction and low contrast explain many errors. A model may also encounter a scene absent from training: a crouching person behind a pallet does not resemble an unobstructed pedestrian.

• Poor angle or a target too small in the image.
• Analysis area including a road, trees or reflective surface.
• Threshold too sensitive or a minimum size inconsistent with perspective.
• Infrared lighting attracting insects and revealing dust or precipitation.
• Changed surroundings after works, vegetation growth or stock movement.

Calibrate on site, then maintain the model

Commissioning tests distances, speeds and clothing by day and night. Exclusions, perspective and schedules are adjusted. Alerts are labelled useful, unwanted, uncertain or missed. Critical tests are repeated after updates, camera movement or seasonal changes. Analytics may run in the camera, on a local server or in the cloud. The choice should reflect latency, network, data location, reversibility, updates and outage behaviour.

A golden rule: no sensitive action on one probability alone

An alert may switch on lighting or display a camera. Accusing someone, imposing a lasting restriction, sharing an image or calling authorities requires context and procedure. The more serious and irreversible the action, the stronger the validation must be. AI prioritises; a person decides.

From signal to response in five steps

A robust procedure defines who receives an alert, target handling time, mandatory checks, contacts and escalation conditions. It also covers ambiguity. If an image is obscured or connectivity fails, the operator should not improvise: degraded-mode instructions say whether to contact someone on site, a patrol or the responsible manager.

• Receive: one located, timestamped incident.
• Qualify: contextual clip, live image and associated sensors.
• Decide: level set by the escalation matrix.
• Act: voice message, lighting, call, intervention or authorities according to the facts.
• Close: record observations and outcome without speculation.

Illustrative scenario: a logistics bay at 02:14

A thermal camera classifies a shape as a person outside a closed warehouse. The visible camera shows someone near the loading bay, but their face is neither necessary nor usable. The system confirms that no credential opened the access and displays the adjacent view. The operator observes a fence crossing, calls the on-duty contact and follows the intervention instruction. This scenario is illustrative: it shows a possible sequence, not a real incident or a promise that the system would prevent intrusion.

Test people and procedures as well

A quarterly exercise can measure handling time, interface comprehension, contact availability and log quality. Harmless mistakes are treated as improvement signals. Security comes from the complete loop: a strong algorithm cannot compensate for an obsolete on-call number, and an excellent team cannot compensate for a camera blinded by a floodlight.

Villas and private property

Cameras cover access points, not neighbours' windows or pavements. Physical angles and privacy masks constrain the view. Person detection may operate around the gate or an outbuilding at appropriate times. A verified alert can trigger lighting, a message or a call. On a large property, an authorised drone or ground robot may provide another angle, although fixed cameras usually remain the most predictable foundation.

Shops, offices and SMEs

Priorities may include out-of-hours entrances, stockrooms, tills or a car park. Staff must be informed, and continuous monitoring of employee behaviour is prohibited in principle. Break areas, workstations and spaces with no security need should not be filmed for convenience. A precise rule such as presence in an armed stockroom is more defensible than general behaviour analysis.

Warehouses, industry and large perimeters

Visible and thermal cameras can cover fencing, vehicle compounds, loading bays and outdoor equipment. Network redundancy, backup power and degraded operation become essential. Cameras work with perimeter sensors, access control and, where useful, mobile systems. A critical site must connect this design to business continuity, cyber-incident response and sector-specific requirements.

Coop: AI-enabled cameras in several Swiss stores

In February 2025, Watson, citing an investigation by Le Temps, reported that Coop used cameras with AI software in several stores, including one in Lausanne, to identify behaviour considered suspicious, notably around self-checkouts. The FDPIC told the publication that any video surveillance had to respect transparency and proportionality. The case documents a current detection capability and the debate it creates; public information does not establish its error rate or effectiveness. For verification, see Video surveillance in the workplace (FDPIC) and Coop uses intelligent cameras in several stores (Watson).

Serco Leisure: biometrics were not necessary

In 2024, the UK Information Commissioner's Office ordered Serco Leisure and associated organisations to stop using face recognition and fingerprints to monitor attendance by more than 2,000 employees at 38 leisure centres. The ICO found that necessity and proportionality had not been demonstrated when cards or badges were less intrusive alternatives. UK law differs from Swiss law, but the design lesson transfers: biometric identification needs far stronger justification than convenience. The applicable guidance is set out in Enforcement against Serco Leisure over employee face recognition (ICO).

SBB: measuring footfall without capturing images

Swiss Federal Railways explains that 3D footfall sensors in stations measure movement, density, dwell time and speed, then transmit only those measurements. They do not capture, record or transmit images and cannot track an individual. This is not security video surveillance. It is valuable because it reverses the question: instead of asking how to film lawfully, first ask whether the desired outcome requires filming at all. The applicable guidance is set out in Passenger flows in railway stations (SBB).

The practical compliance file

Before procurement, the controller describes the purpose, zones, affected people, schedules, data categories, recipients, storage and retention. Less intrusive alternatives are documented. A notice should be visible before someone enters the field of view and identify the controller. Tested procedures should cover access requests, deletion, extracts for authorities and data breaches.

• Limit each camera's field physically and digitally.
• Disable microphones unless strictly necessary.
• Mask zones, faces or feeds where identification is unnecessary.
• Use named permissions and log viewing.
• Set automatic deletion with a controlled hold after an incident.
• Assess overseas transfers and cloud processors.

Employees require greater care

Employment law adds its own limits. The FDPIC says systems intended to monitor employee behaviour are prohibited and continuous surveillance may harm health. Security cameras may be possible at selected entrances, car parks, halls, hazardous areas or valuable stock where a less intrusive measure cannot meet the purpose. Inform employees before commissioning and keep their appearance in frame as exceptional as possible.

Lawful does not automatically mean acceptable

A system may meet formal duties yet remain poorly understood. Explaining what is and is not detected, who watches, how long data exists and how an error can be challenged builds trust and often reveals unnecessary settings. Transparency is an operational-quality measure: a purpose that cannot be explained simply is probably still too broad.

The minimum technical baseline

Place cameras on a dedicated network separated from office systems and industrial controls. Disable unused services and manage devices through controlled paths, ideally with strong authentication. Remove generic accounts. Send access and configuration logs to monitoring. Encrypt configuration backups and test restoration.

The contract should say who fixes vulnerabilities, within what time, for how many years and how third-party components are governed. Secure boot and signed software strengthen the trust chain but do not replace updates or hardening. Supplier departure, camera replacement and end of life should include access revocation, data erasure and evidence that the asset left the inventory.

• Unique identity and separate secret for every device.
• Dedicated VLAN or segment with only required network flows.
• Encrypted streams and lifecycle-managed certificates.
• Patches, vulnerability monitoring and a support deadline.
• Tests of restoration, network loss and backup power.
• No direct internet access to camera administration.

A four-decision roadmap

First, prioritise three to five incident scenarios and their actions. Second, test sensors on site against measurable criteria: coverage, useful-alert rate, missed events, verification time and availability. Third, approve the legal, cyber and organisational file before full deployment. Fourth, operate through regular reviews of alerts, site changes, software versions, permissions and exercises.

The useful return indicator is not camera count. Combine avoided patrols, verification speed, evidence quality, system availability and genuinely actionable alerts. A modular architecture can later add a robot, drone or new analytics when the need is demonstrated, without rebuilding the entire chain.