\"SecureSafety PPE detection — the result when camera infrastructure passes a readiness survey.\"
The operations director had done everything that the procurement process asked of him. He had selected the vendor with the strongest case studies. He had secured the capital budget. He had arranged a phased deployment across twelve cameras covering the vehicle marshalling yard, which was the site's highest-priority area for pedestrian-vehicle incidents.
What nobody had checked — because nobody had framed the question — was whether the cameras that the AI would run on were actually capable of supporting it.
Four were analogue cameras feeding through a digitiser that capped the output at 240p effective resolution. Three used a proprietary compression codec that the edge AI device could not decode natively, requiring transcoding that added 1.4 seconds of latency to the detection pipeline. Two were positioned with the afternoon sun directly behind them from about 14:30 onwards in winter — which happened to coincide with shift change. The remaining three performed acceptably.
Three acceptable cameras on a twelve-camera deployment is not a pilot. It is a near-complete waste of budget, a frustrating data point for the board, and a missed opportunity to protect a hazardous area during the window when it matters most.
This is not an unusual story. It is the predictable outcome of a sales process that glosses over the camera infrastructure question and a buyer who did not know to ask it.
Why camera infrastructure matters more than the AI model
Computer vision AI processes the image stream it receives. It cannot compensate for information that was lost before the stream reached it. A 4K detection model receiving a 240p video feed will not produce 4K detection results; it will produce results that reflect the quality of the 240p feed — with all the limitations that entails for identifying a specific PPE item at seven metres, or distinguishing a pedestrian from a column shadow in low light.
Three factors most consistently constrain detection quality in real industrial camera estates:
Resolution at working distance determines how much visual information is available at the point where detection needs to happen. The practical measure is not the camera's headline megapixel rating but the effective pixel count covering the detection target — a person's head and torso, or a vehicle — at the maximum relevant working distance. The rule of thumb most widely used by integrators is a minimum of 40 pixels covering the key target feature at the furthest relevant distance. Below that threshold, even a well-trained detection model becomes unreliable, because the information simply is not in the image.
Frame rate determines whether fast-moving events are captured between frames. A forklift travelling at walking speed — 5 km/h — moves approximately 1.4 metres per second. At 1 frame per second, the vehicle can traverse 1.4 metres between consecutive frames, potentially passing through a detection zone without triggering a proximity alert. Vehicle-pedestrian detection requires a minimum of 10 frames per second, with 15–25 fps being the practical operating range for reliable real-time detection.
Compression and bitrate is the least-discussed and most consistently underestimated variable. H.264 and H.265 compression introduce artefacts — blocking, blurring, ghosting from previous frames — that compound in older NVR equipment or in streams that have been compressed, transmitted across a congested network, decompressed, and recompressed. A camera's headline specification means less than its effective bitrate in operation: a 4MP camera configured at 256 kbps produces worse detection-quality images than a 2MP camera running at 2 Mbps.
Beyond these three core factors, angle, field of view, environmental conditions, and network infrastructure all introduce further variables that determine whether a detection system can perform its designed function.
The 12-point camera survey
Run the following survey for every camera in the candidate coverage area before finalising deployment scope or budget. Score each criterion per camera: 3 = meets requirement fully; 2 = marginal, may require configuration change; 1 = below threshold, requires hardware or infrastructure change; 0 = fundamentally unsuitable.
The survey covers every installed camera's resolution, frame rate, field of view, compression, IR mode, and VMS compatibility — before any AI configuration begins.
1. Resolution at target distance What is the effective pixel count covering the detection target — worker's head and torso, or vehicle — at the maximum working distance from this camera? Minimum acceptable: 40 pixels on the key target feature. Verify with a tape measure, a test walk at maximum distance, and a frame grab reviewed at native resolution. Do not rely on the camera's label specification; measure it.
2. Frame rate What is the configured frame rate on the camera stream as it reaches the NVR? Minimum for pedestrian monitoring: 10 fps. Minimum for vehicle-speed monitoring: 15 fps. Check the NVR configuration directly — many cameras are installed at a frame rate significantly below their maximum capability as a storage-saving measure.
3. Compression and bitrate What is the configured bitrate for this camera's stream? Request a frame grab at different times of day and inspect it at 100% zoom for compression blocking artefacts. Minimum: 1 Mbps for a 2MP camera; 2 Mbps for a 4MP camera. Severe artefacting at existing bitrate typically requires NVR configuration changes, though older NVRs may not support higher bitrates without hardware upgrade.
4. Night mode and IR capability Does this camera operate in IR night mode? What is the effective illuminated range in IR mode? Is the IR illumination adequate for the full detection area after dark? Cameras with IR range below the working distance of the detection zone leave portions of the coverage area in effective darkness after the illumination threshold.
5. Field of view and occlusion Does the current field of view cover the complete area of interest? Are there permanent occlusions — structural columns, racking, large fixed equipment, signage — that create blind spots within the monitored zone? Create a simple coverage diagram for each camera, marking the field of view and annotating any occlusions. Occlusions that cover the highest-risk area (the closest pedestrian–vehicle interaction point) should be flagged as critical.
6. Camera angle to detection target What is the angle of incidence from camera lens to the detection target? Overhead cameras at greater than 60 degrees elevation are effective for monitoring overhead hazards and general counting but are poor for PPE detection (hard hat versus bare head requires a near-frontal view) and face-recognition quality. Cameras at 15–45 degrees elevation provide the best geometry for pedestrian PPE detection and identity. Record the angle for each detection use case the camera is expected to serve.
7. VMS protocol and codec compatibility What is the camera's RTSP stream address? What codec does it output — H.264, H.265, MJPEG, a proprietary format? Is that codec natively supported by the edge AI device, or does it require transcoding? Codec incompatibility is the single most common technical failure point in retrofit deployments. A camera that produces a proprietary stream requires a transcoding step that adds latency and introduces additional compression artefacts. Confirm native codec support before including a camera in the deployment scope.
8. Network bandwidth What is the available network bandwidth on the path between each camera and the edge AI device? For multiple cameras feeding a single edge device, the aggregate bandwidth requirement may exceed the capacity of the existing switch port or network segment. Measure actual sustained throughput, not theoretical rated capacity — they frequently differ in legacy industrial network installations.
9. Camera maintenance status When was this camera last serviced? Is the housing intact and weatherproof? Are there signs of water ingress, cable abrasion, corrosion at connectors, or physical impact damage to the housing or mounting bracket? A camera with degraded weatherproofing in a UK outdoor environment will deteriorate further; performance in winter will be significantly worse than at the time of survey.
10. Lens cleanliness Is the lens clean and free of contamination? IR illuminator windows, in particular, accumulate spider webs, dust, condensation deposits, and airborne contamination from industrial processes. In external environments, lens surface degradation is one of the most common — and most under-reported — causes of night-mode detection failure. This criterion sounds trivial; it is not.
11. NVR and cabling condition What is the age and condition of the NVR receiving this camera's feed? Does the NVR introduce additional compression or buffering latency? Is the cabling — Cat5e, coaxial, or fibre — in good condition with no known faults? Old coaxial runs in particular are a source of signal quality degradation that is invisible in the cable's physical appearance but material in the image quality it delivers.
12. PTZ configuration (if applicable) Is this a pan-tilt-zoom camera? PTZ cameras can only cover the area where their lens is currently pointed. AI monitoring requires a fixed field of view to provide consistent, always-on coverage of a defined zone. PTZ cameras may be usable for safety AI if a fixed "safety zone" preset is configured and PTZ automation is locked during monitoring hours — but this must be explicitly confirmed in writing, because PTZ automation that overrides the preset will create detection gaps.
Scoring and interpretation
Add the scores across all 12 criteria for each camera:
| Total Score | Assessment | Recommended Action |
|---|---|---|
| 30–36 | Fully suitable | Proceed with deployment at this camera |
| 22–29 | Suitable with configuration changes | Agree a specific remediation plan before deployment; proceed once complete |
| 15–21 | Marginal — hardware or infrastructure investment required | Budget for remediation; do not include in pilot scope until complete |
| Below 15 | Not suitable in current state | Replace camera, add supplementary camera, or accept that this location cannot be covered |
Cameras scoring between 22 and 29 almost always require one or two targeted configuration changes — bitrate increase on the NVR, fixed preset configuration on a PTZ camera, IR illuminator cleaning — that cost little and can be completed before deployment begins. Cameras scoring below 15 typically have a structural problem (inadequate resolution, incorrect geometry, codec incompatibility) that configuration changes cannot resolve.
What to fix versus what to accept
Not every deficiency requires resolution before a deployment can proceed. The right frame is: does this deficiency affect detection at the highest-priority hazard zone on this camera?
Fix before deployment starts: codec incompatibility (adds latency and artefacts that no algorithm compensates for); RTSP stream inaccessibility (no stream, no deployment); frame rate below 10 fps on vehicle-coverage zones; critical angle deficiencies at the highest-priority detection points.
Fix during the pilot period: bitrate configuration below minimum (usually an NVR setting change, low cost and zero hardware change); lens and IR window cleanliness; minor field-of-view adjustments achievable by remounting the camera.
Accept with a documented limitation: PTZ cameras in a locked fixed-preset configuration, provided the locked view covers the critical detection area and the lock is verified on a weekly basis; modest resolution deficiencies at secondary cameras where primary coverage of the same area is adequate; minor occlusions that do not include the highest-risk point.
Replace or supplement: cameras below 30-metre effective IR range on zones with vehicle-speed detection requirements; analogue cameras with digitisers delivering below 720p equivalent resolution at working distance; cameras with angle problems that prevent detection of the relevant target class at the relevant geometry, where remounting is not feasible.
When to add cameras rather than adapt to limitations
The strongest case for adding a camera rather than working around existing limitations arises when a high-priority detection need cannot be adequately served by any existing camera, regardless of configuration changes. A restricted zone at a pedestrian-vehicle crossing point with no camera providing an adequate angle within practical cable run warrants a new camera. The cost of a single modern IP camera and its installation — typically £500–£1,500 all-in for an onshore industrial installation — is modest relative to the cost of the incident that goes undetected because the coverage gap was not addressed.
The BSI standard BS EN 62676 (CCTV systems for use in security applications) provides the baseline framework for camera performance specification in industrial settings, including the definition of the detection, observation, recognition, and identification quality tiers that correspond to different minimum pixel-on-target requirements. While developed for security applications, the specifications translate directly to safety AI use cases. The National Protective Security Authority (NPSA) and the National Security Inspectorate (NSI) publish supplementary guidance on camera placement and specification for high-security environments that is equally applicable to safety-critical coverage zones.
The oil and gas platforms where SecureSafety first deployed safety AI presented camera infrastructure challenges that make most onshore industrial estates appear straightforward: salt-laden air, temperature cycling across 55 degrees, constant vibration from drilling equipment, camera cable runs of 200 metres in explosive-atmosphere rated conduit, and no possibility of a site visit to swap a camera that underperforms after installation. Building a detection system that held a sub-0.05% error rate under those conditions required understanding camera infrastructure at a level of specificity that most software vendors never develop. The 12-point survey above reflects that experience. When SecureSafety conducts a Discovery engagement for a new site, the camera infrastructure assessment is the first deliverable — because everything else depends on it. Book a demo and we will run this survey across your site's camera estate before you commit a penny to a deployment.
