At night, the AI alert pipeline does not just turn on the IR light. Three inputs change weight.

By treating night as a different operating mode, not as the day pipeline with the IR illuminator turned on. A working night-mode classifier reweights three inputs at sunset. (1) The time window itself becomes a primary HIGH-threat trigger for after-hours zones, because at 03:00 a person at the rear gate is not plausibly a delivery driver or a tenant returning from work. (2) Posture inference is downweighted, because IR mode collapses the scene to grayscale, kills clothing pattern, and pushes the posture model below the confidence floor for most frames. (3) The environmental-motion silencer flips its noise model from day sources (foliage, shadow shifts, sun glare) to night sources (insects on the IR lens, headlight glare from passing vehicles, IR hot-spot reflection on glass, and the half-second of color-IR transition when the IR-cut filter switches). The result on a typical multifamily property is about 30 to 60 raw detections per camera per night collapsing to 1 to 3 paged events.

Most posture classifiers were trained on color RGB crops. Under IR, the sensor switches to a monochrome path and the IR illuminator floods the scene with near-infrared light at roughly 850 nanometers. Three things degrade at once. Clothing patterns vanish, because dyes that look distinct in visible light reflect near-IR similarly. Skin tone disappears, so head and hand boundaries get fuzzier. And the IR illuminator itself produces a hot spot in the center of the frame and a fall-off at the edges, which warps the texture distribution the model trained on. The posture model still emits a class, but its confidence on poses like 'leaning' versus 'walking' versus 'hands-on-door' drops from a comfortable 0.85 day-time average to under 0.6 in IR. Below that floor, the right behavior is to cut the posture weight in half and let zone, time, dwell, and badge state carry more of the decision.

Four matter in practice. (1) Insects landing on or near the IR lens. The IR illuminator attracts moths and mosquitoes, which then fly into the field of view and trigger a person-or-vehicle classifier briefly before the bounding box dissolves. Day cameras do not see this because no IR illuminator is running. (2) Headlight glare from passing vehicles. A car on the perimeter road at midnight produces a moving column of high-luminance pixels that crosses several zones in 1 to 2 seconds. The track stitcher needs to recognize this as a vehicle and not as a fast-moving person. (3) IR hot-spot reflection. When the illuminator is mounted close to a window, a glass door, or a metal mailbox, the reflection produces a bright stationary blob that the detector occasionally promotes to a person at low confidence. (4) The IR-cut filter switching itself. At dusk and dawn, the camera slides a mechanical filter into place to switch between color and monochrome modes, and during that 200 to 800 millisecond transition the whole frame goes through an exposure swing that produces a wave of motion that no real object generated.

For after-hours-restricted zones, yes. The intent classifier's day rule for a person at the rear gate is 'unknown badge plus dwell over 60 seconds equals HIGH'. The night rule for the same zone is 'unknown badge plus any dwell equals HIGH', because at 03:00 there is no benign reason for an unknown person to be at the rear gate at all. For zones that remain semi-public at night (lobby, parking) the time window does not auto-trigger but it does lower the dwell threshold, typically by 30 to 50 percent. The classifier still requires class plus zone plus dwell to fire, but the dwell bar is lower so the alert lands faster. The reason this works is that the operator is asleep and false-HIGH at 03:00 costs them about ten times more than a false-HIGH at noon. Tighter time-based gating recovers signal at a lower volume.

A false-HIGH at noon interrupts a meeting. A false-HIGH at 03:00 wakes the on-call manager, costs them 20 to 40 minutes of sleep latency on the recovery, and trains them to silence the channel by week three. The cost of a missed real event is the same at any hour, but the cost of a false event is roughly an order of magnitude higher overnight. Tuning has to reflect that. Practical numbers we hold to on production sites: target under 0.3 false-HIGH per camera per night, versus under 1.0 per camera per week in daylight. The two ways we get there are the time-window reweighting above and a stricter posture override (a 'hands-on-door at 02:00' alert requires confirming dwell over 8 seconds before paging, even though the same posture during the day pages immediately).

The IR-cut filter is a small mechanical filter that physically slides in front of the sensor. During the day it blocks IR; at night it slides out so the sensor can see IR. The transition is triggered by an ambient light sensor or a luminance threshold on the frame itself, typically firing once at sunset and once at sunrise per camera, with occasional thrash on cloudy dawns. While the filter is moving, the sensor's gain auto-adjusts and the entire frame goes through a 200 to 800 millisecond exposure swing. During that swing, every part of the frame appears to be moving, and a naive motion-difference engine fires on the whole composite. A working classifier suppresses this in two ways. First, the edge unit detects the bimodal luminance histogram shift that signals a filter swap and gates classification for one second after the event. Second, the per-frame detector's confidence floor is raised to 0.85 during that gate, so any survivor classification has to be unusually clean. The recorder's own motion engine has neither of these and will fire a wave of false events at sunset every single day.

They are gone. The IR sensor path is monochrome. A search query like 'person in red shirt at the rear gate' resolves zero hits at night because the camera literally did not record color. What survives at night and is still searchable: shape, posture, gait, vehicle silhouette, license plate (if there is a dedicated LPR camera in the cluster), badge state from access control, and zone-and-time. A night-aware search index expands the query at runtime, so 'person in red shirt at 02:30' becomes 'person at 02:30, color filter dropped' and returns the matching detection without crashing on the missing color feature. This matters when an operator is reviewing footage at 09:00 the next morning trying to find a specific person and types in a daylight description.

On the inference side, no. The detection-to-classification path runs at the same composite frame rate whether the camera is in color or IR mode, because the model accepts whatever the sensor emits. On the messaging side, the page itself is the same SMS plus voice call to on-call. Where night does change is the operator's ack latency. Day acks land in 10 to 30 seconds. Night acks land in 60 to 180 seconds because the manager has to wake up, find the phone, and pull the clip. We compensate for that by attaching the 10-second triggering clip to the SMS itself, so the manager can decide whether to escalate before fully sitting up. The practical effect is that the action window at night is wider by about a minute, but it starts deeper in the incident timeline.

Yes, and night is actually the easier case to retrofit. The reason: the recorder is already drawing every camera onto a multiview composite for the wall display, and it is already handling the IR-cut filter and the IR illuminator at the camera level. An edge unit reads the recorder's HDMI output, splits it into per-tile crops, runs detection on each tile, and applies the night-mode reweighting in software. None of the cameras change, none of the IR illuminators change, and the recorder keeps its retention schedule. The edge unit handles up to 25 feeds, runs entirely on-device (no cloud round trip for alerts), installs in roughly two minutes (one HDMI cable from the recorder out), and adds the night-mode classifier on top of the detection layer the box was going to run anyway. The production unit we ship is $450 one-time and $200 per month from month two.

Numbers we see on production deployments at Class B and C multifamily, 8 to 15 cameras per property, IR illumination at every exterior camera. Raw detections from the per-frame detector: roughly 30 to 60 per camera per night, dominated by insects and headlight glare. After the night-mode classifier (zone, time, dwell, badge, downweighted posture, environmental silencer): 1 to 3 paged events per camera per night, of which about 1 in 10 is a real intrusion or trespass. The remaining 9 in 10 are LOW THREAT events that go to the digest. False-HIGH rate: under 0.3 per camera per night. False-LOW rate (real event missed at HIGH): under 1 per camera per quarter, audited weekly by reviewing the digest for events that should have paged. At a 180-unit Class C property in Fort Worth, this configuration caught 20 incidents including a break-in attempt in the first 30 days; the customer renewed at the end of month one.