A helmet detector that scores well at 2pm in an open yard can fall over at 11pm at the same gate, on the same camera, with no change to the software. The cause is almost always the image rather than the model: fewer photons per pixel, a longer exposure, a monochrome infrared frame, or a lens coated in cement dust. ARSA publishes the thresholds that separate the two cases in ARSA-CVS-001, and this article explains what those numbers mean on a night shift.
THE GOVERNING CONSTRAINT
Software Cannot Recover Light The Sensor Never Collected
PPE classification is a high pixel-density task. A helmet has to be told apart from a cap, a bandana and a bare head, and that distinction lives in the shape of the brim and the shell. ARSA-CVS-001 sets 262 pixels per metre for helmet presence and 153 PPM for hi-vis vest presence. Pixel density falls as one over distance, so a 2 MP camera on a 60 degree lens reaches 6.3 m for helmets and 10.9 m for vests, and no amount of tuning extends that.
Light enters the same equation from a different direction. The standard requires a minimum of 50 lux at the target plane for both helmet and vest detection. For comparison, EN 12464-1 puts coarse assembly work at 300 lux and warehouse gangways at around 150 lux, so a compliant working area is usually several times above the analytic floor. The problem areas are the ones nobody lights for work: a night gate, a perimeter road, a loading bay lit only by a single wall pack, a stockpile at a mine site. Those are the positions that fail, and they fail quietly, by producing fewer detections rather than obviously wrong ones.
WHAT ACTUALLY DEGRADES
Three Ways A Dark Scene Defeats A Working Camera
Automatic Shutter Smears The Target
Most cameras ship tuned for a human watching a monitor. As light drops, automatic exposure lengthens the shutter to keep the picture bright. A walking worker then arrives at the sensor as a blurred column, and the helmet brim that the 262 PPM requirement exists to resolve is spread across pixels that no longer carry an edge. The fix is a manually pinned minimum shutter on the analytic stream, set per analytic, accepting a darker frame in exchange for a sharp one.
Infrared Removes Colour Entirely
Infrared illumination at 850 nm produces a monochrome image. Helmet detection survives that, because a helmet is a shape. Hi-vis vest detection does not, because a vest is defined by colour and retroreflective response, and every colour-coded zone rule goes with it. Night operation of vest detection requires white light at the target plane. This is a hard boundary rather than a tuning preference, and it is worth deciding at design time whether a given gate needs vest coverage after dark, because the answer changes the lighting budget.
Smart Codecs Delete The Evidence
H.264+, H.265+ and equivalent bitrate-saving modes compress the static background hard and allocate bits to motion. On a dim scene with sensor noise, the encoder reads noise as activity and starves the region where a small distant target needs detail. Variable bitrate compounds it by throttling exactly when a shift change fills the frame. Constant bitrate with smart codec disabled on the analytic stream costs storage and returns the pixels the detector needs.
RESOLUTION
More Megapixels Is The Wrong Reflex At Night
The instinct when accuracy drops is to buy a higher-resolution camera. Each step up divides the light collected by every pixel unless the sensor grows physically larger. An 8 MP sensor of the same format collects roughly a quarter of the light per pixel that a 2 MP sensor does, and carries about 3.5 times the bitrate and storage for twice the reach. In a dim area, a 4 MP camera with a longer focal length usually beats an 8 MP camera with a wide lens, because the narrower lens roughly doubles pixel density while halving the covered width. The lens and field of view visualiser shows that trade for a specific mount before anything is purchased.
THE NUMBERS
Thresholds For PPE Detection In Difficult Conditions
| Parameter | Helmet Detection | Vest Detection |
|---|---|---|
| Pixel density at target | 262 PPM | 153 PPM |
| Maximum range, 2 MP, 60° lens | 6.3 m | 10.9 m |
| Maximum range, 4 MP, 60° lens | 8.4 m | 14.5 m |
| Optimal tilt below horizontal | 10° to 25° | 10° to 25° |
| Maximum tilt | 35° | 40° |
| Tilt steeper than 45° | Prohibited | Prohibited |
| Minimum light at target | 50 lux | 50 lux |
| Infrared illumination only | Supported | Unsupported |
| Site And Camera Parameter | Minimum | Recommended |
|---|---|---|
| Sensor resolution | 2 MP | 4 MP |
| Frame rate, main stream | 15 fps | 25 fps |
| Low-light sensitivity, colour | ≤ 0.05 lux @F1.6 | ≤ 0.01 lux @F1.4 |
| Wide dynamic range | Not required | ≥ 120 dB |
| Ingress protection, outdoor | IP66 | IP66 with IK08 |
| Operating temperature | −10 to +50 °C | −20 to +60 °C |
| Mount rigidity | Wall or soffit | Pole with damping |
The accepted reference case in ARSA-CVS-001 is a 3.5 m mount at 6.0 m standoff with 22 degrees of tilt, which presents helmet, vest and torso frontally and meets 262 PPM at the far gate line. Tilt matters as much as light. A camera looking steeply down sees the crown of a helmet and little else, because the wearer’s own body occludes the vest front and any harness strap. Above 45 degrees, PPE analytics are prohibited outright.
DUST, HEAT AND VIBRATION
Extreme Sites Fail On The Enclosure Before The Algorithm
In a crusher house, a cement plant or an open-pit haul road, the limiting factor is usually the housing. IP66 under IEC 60529 means dust-tight and protected against powerful water jets, which is the floor for any outdoor or high-particulate position. IK08 under IEC 62262 corresponds to a 5 joule impact, which is what stops a thrown tool or flying debris from ending coverage of a zone.
Dust affects accuracy long before it affects uptime. A film on the dome scatters light, flattens contrast, and drops effective pixel density well below the figure a design review calculated from geometry alone. Airborne dust between camera and target does the same in the air column. Both are cleaning and mounting problems rather than model problems, and both are why a position that passed at commissioning should be re-surveyed when conditions change.
Vibration is the third one. A pole-mounted camera near reciprocating plant oscillates, and a long lens amplifies that oscillation. Pole mounts need damping. Power deserves attention too, since infrared illuminators and enclosure heaters peak together in the wet season, so the PoE budget should carry at least 25 percent headroom.
WHAT THE SYSTEM DOES NOT DO
Stated Limits
The Basic Safety Guard detects helmet, vest and boot presence or absence, hazard-zone entry, and supports custom add-on items. It does not verify that a helmet is correctly fastened, does not read a vest’s certification class, and does not grade the condition of equipment. It does not work through infrared for any colour-dependent rule. It does not produce an accuracy commitment for a zone that has never been surveyed: ARSA warrants figures only for detection zones recorded as ACCEPTED in the Site Acceptance Record, and a zone reverts to unaccepted status when a camera is moved, lighting changes, or racking obstructs the view. Detection runs on the appliance on site using standard deep learning frameworks, and video leaves the network only where the operator explicitly configures it.
FREQUENTLY ASKED
Questions Engineers Ask Before Commissioning
Does PPE Detection Work In The Dark?
Helmet detection works down to 50 lux at the target plane. Below that, the camera lengthens its shutter or raises gain, and the result is blur or noise rather than a clean detection. Vest detection needs 50 lux of white light, since infrared removes the colour the class depends on.
Can Infrared Illumination Replace Site Lighting?
For helmet presence and zone intrusion, infrared at 850 nm is usable. For hi-vis vests and any colour-coded rule, it is not, because the image is monochrome. A site that needs vest enforcement at night needs white light at the target plane.
Will Dust Or Rain Stop Detection?
Heavy airborne dust, fog and rain reduce contrast and effective pixel density, so accuracy degrades progressively rather than switching off. An IP66 housing with regular dome cleaning holds most of the loss. Where a scene is persistently obscured, the honest answer is a closer camera rather than a tuned model.
Our Plant Runs At 55 °C Near The Furnace. Is That Within Spec?
The contracted envelope is −10 to +50 °C, with −20 to +60 °C recommended hardware. A position above that needs a cooled or ventilated enclosure, or a mount further from the heat source with a longer lens to hold pixel density.
How Do We Find Out Which Positions Fail Before We Buy?
A Remote Camera Design Review is $1,500 and returns a pass or fail against every requirement in ARSA-CVS-001 for up to 20 cameras within one week. An On-Site Camera Survey is $4,500 plus travel for up to 50 cameras. Either fee is deducted in full from the project fee if you contract within 90 days.
What Happens To A Camera That Fails?
The review states the specific remediation and prices it: a different lens, a different position, added lighting, or an extra camera. You decide what to remediate, and unremediated zones sit outside the accuracy commitment.
NEXT STEP
Check The Positions Before You Commit To The Software
Low-light and extreme-environment performance is decided by geometry, illumination and enclosure choices that are cheap to change on paper and expensive to change on a pole. Read the published thresholds on system requirements, then look at Basic Safety Guard to see what runs on the cameras that pass.
Sources:
– How Many Lux for Industrial Lighting? Complete Guide (EN 12464-1)
– IEC 60529 Ingress Protection (IP Code) Certification Testing
– IP vs. IK Ratings: Water, Dust, and Impact Protection