Definition, scope, and product boundaries
A vision-based sleep monitor uses optical imaging components to acquire two-dimensional images, depth maps, or thermal distributions of a sleep space continuously or intermittently. It then applies person detection, motion analysis, or physiological-estimation algorithms to produce information such as bed presence, posture, body movement, and respiration. Within this encyclopedia it belongs to WSIC 1.2.1, Vital-sign Monitoring Devices. “Vision” here means an imaging chain with spatial pixels or a distance array, not simply a person watching recorded video.
Visible-light RGB, near-infrared night vision, structured-light or time-of-flight depth, stereoscopic vision, and thermal imaging can all be vision modalities, but they measure different quantities under different conditions. Infrared night-vision sleep-monitoring cameras and infrared thermal-imaging sleep monitors are specific branches of this category. Products whose principal input is radio reflection, sound, mattress pressure, or a body-worn sensor should instead be classified by that primary measurement principle.
A network camera with recording and motion detection is not thereby a sleep monitor. A finished system also needs a specified field of view, night imaging, timing, person-to-bed association, quality control, event or physiology algorithms, and verifiable outputs. A camera used as the observational channel during polysomnography is not the same use as a system that estimates sleep metrics from video alone. The former lets staff review behaviour; the latter also requires independent validation of its automated algorithms.
Imaging modalities, hardware, and data chain
An RGB camera records reflected visible light and may analyse outline, posture, movement, and small colour changes in visible skin, but it depends on illumination and exposure. A near-infrared system combines infrared illumination with a suitable sensor for dark conditions. A depth camera outputs scene distance or surface shape, while a thermal camera records radiance in a defined spectral band. A greyscale night image, a depth value, and a temperature-related pixel are not interchangeable physiological quantities, and display colour alone does not identify the sensor type.
A typical system consists of a lens, image sensor, optional illuminator or depth emitter, clock, edge processor, storage, network interface, firmware, and application. Important parameters include field of view, working distance, pixel and depth resolution, frame rate, exposure, dynamic range, illumination wavelength and power, timestamp accuracy, and edge-compute capacity. A high nominal resolution does not guarantee that small motion or skin signals remain adequate after bed-distance imaging, low illumination, and compressed transmission.
Processing commonly starts by locating the bed and person and then performs foreground segmentation, key-point or contour extraction, target tracking, and occlusion assessment. It may next calculate optical flow, surface displacement, colour variation, or thermal change before producing events and trends. Cropping, down-sampling, compression, automatic exposure, variable frame rate, and model preprocessing all alter information. The system should disclose whether original frames exist, which regions are used, when quality is inadequate, and which model version ran instead of supplying only an untraceable composite score.
Posture, movement, and physiological outputs
Relatively direct visual outputs include empty or occupied bed, bedside activity, bed exit, larger movement, turning periods, and outline-based categories such as supine, prone, or lateral posture. Head and neck direction, limb placement, and fine postures beneath bedding are more vulnerable to occlusion, so coarse and fine classifications require separate validation. An algorithmic “still” label means little visible change in the analysed region; it does not by itself establish sleep, absence of pain, or absence of quiet wakefulness.
Respiration may be estimated indirectly from periodic displacement of the chest, abdomen, or bedding, or from thermal change near the nose and mouth. Pulse-related information may be inferred from small periodic variation in light reflected by visible skin. These routes have different requirements for exposed skin, illumination, camera distance, movement, and bedding. Video does not directly measure airflow, respiratory effort, electrocardiography, or oxygen saturation. Research estimates of camera-based oxygen saturation, blood pressure, or apnoeic events cannot be generalized without the actual wavelengths, reference method, and target population.
Sleep onset, awakenings, total sleep time, and sleep stages are interpretive outputs in which a model further combines movement, respiratory, pulse, and temporal features. Clinical staging relies on signals such as electroencephalography, eye movement, and muscle activity. Video-model labels are interpretable only within a stated scope after epoch-by-epoch validation against synchronized polysomnography. Posture and respiratory trends can help describe a night, but they cannot independently establish obstructive events, parasomnia, or another sleep disorder.
Installation, occlusion, and environmental failure
Camera position, height, downward angle, focal length, and field of view should be selected for bed width, number of occupants, sleep positions, and intended output. The lens must be fixed and its position marked. Acceptance testing should use the actual bed and night conditions and include empty bed, bed entry, common postures, turning, sitting up, and bed exit. Moving the camera, bed, or furniture, or changing image crop, bed-head elevation, curtains, or lighting can invalidate the previous bed region and scale and requires renewed checking.
A thick duvet, pillow, bed rail, mosquito net, caregiver, bed partner, or pet can obscure the person or confuse target association. Deformation of loose bedding can also be mistaken for chest or abdominal motion. Darkness, backlighting, flicker, infrared reflection, lens dirt, condensation, automatic exposure, and video compression affect image quality. Thermal imaging additionally depends on surface temperature, environmental heat sources, and bedding insulation and cannot see through covers. A depth camera may be affected by range, reflective material, or another active optical source.
With two people in a bed, posture ownership, physiological-signal separation, and the identity of the person leaving must each be validated. A single-person model does not become a multi-person monitor merely by widening the frame. Defocus, occlusion, leaving the frame, insufficient light, dropped frames, and uncertain identity should be labelled invalid or low confidence rather than filled with normal-looking traces. Loss of network or power, full storage, clock drift, and cloud outage should also be exercised before deployment, with local operation and recovery rules documented.
Performance validation and procurement acceptance
Validation should be designed separately for each output. Person or bed detection requires missed and false detections and identity switches. Posture classification requires category definitions, a confusion matrix, and separate coarse- and fine-grained results. Bed-exit or unusual-motion events require event sensitivity, positive predictive value, false alarms per night, and notification latency. Continuous quantities such as respiratory or pulse rate require valid coverage, error, bias, limits of agreement, and recovery time after movement. One overall “accuracy” cannot replace these measures.
Sleep analysis should be compared epoch by epoch with a suitably synchronized reference and should report sleep and wake sensitivity and specificity, stage performance, all-night errors such as total sleep time, and the proportion of invalid nights. A study should also describe people and nights, age and health, skin tone and visible-skin conditions, bedding, illumination, camera distance, multiple occupancy, reference annotation, exclusions, and hardware, firmware, and model versions. Removing occluded or failed nights before analysis overstates real deployment performance.
Procurement should include a multi-night parallel pilot in the intended rooms, with acceptance thresholds, authority to retain raw evidence, an invalid-data limit, a privacy method, and alarm ownership defined in advance. Evidence must apply to the complete product and actual mode. A development camera, short video clip, vendor dataset, or predecessor algorithm cannot replace validation of the current model. Changes to camera settings, compression, or the model should be versioned and followed by risk-proportionate regression checks, especially for occlusion, darkness, different skin tones, and multiple people.
Image privacy, network, and physical safety
Bedroom imagery can disclose identity, the body, routines, household relationships, care activity, and the home environment. It remains highly sensitive even when collected for sleep analysis. Before activation, everyone who may enter the field of view should be told the imaged area, whether audio is captured, where processing occurs, how long data are retained, who receives them, and under what conditions they can be viewed. Practical controls should support pausing, physically covering the lens, deletion, and export. Institutional use also requires assigned responsibility and audit for live viewing, event review, and emergency exceptions.
Edge processing that outputs only an outline, skeleton, feature, or event can reduce transmission of raw images, but it does not automatically anonymize the data. Reversible imagery, facial templates, key points, heat maps, and timestamps may still be linkable to a person. The system needs device identity, unique credentials, least privilege, protection in transit and at rest, access logs, signed updates, vulnerability response, and secure deletion at end of service. It should state what is retained by cloud, mobile client, third-party analytics, and backups. Nonessential microphones and continuous raw recording should be off or minimized by default.
“Invisible” near-infrared or structured light still requires photobiological-safety assessment. An illuminator should meet applicable requirements for its actual wavelength, power, distance, and operating mode; a generic LED certificate does not replace evaluation of the assembled product. Mounts must be secure, and devices and cables must not fall, entangle a person, or obstruct bed exit and care. Indicator light, fill illumination, and fan noise should not disturb sleep unnecessarily. Network, electrical, or optical safety conformity does not establish physiological accuracy of the algorithm.
Use, maintenance, medical limits, and response
Routine checks should cover camera fixing, field of view, cleanliness, focus, illumination, clock, network, storage capacity, and the alarm path. The lens and enclosure should be cleaned only as instructed. An object blocking the lens or a change in layout, bedding, occupants, firmware, or model should be recorded and followed by repeat testing. Recurrent blanks, image discontinuities, implausibly smooth traces, or events that contradict observation should prompt investigation of imaging and data flow before being interpreted as a change in the person.
A general-wellness vision system may help review posture, movement, and trends and may provide material for a professional discussion, but it cannot replace polysomnography, a validated home sleep-apnoea test, or clinical assessment. A product claiming diagnosis, screening, continuous medical surveillance, or treatment-triggering action requires the applicable medical-device authorization, target population, contraindications, performance evidence, and human-review workflow in the market of sale. “Artificial intelligence,” “contactless,” and “derived from hospital technology” do not lower these requirements.
A normal display does not rule out real symptoms, while an alert should not lead to self-treatment without symptoms and professional judgment being considered. Witnessed breathing pauses, awakening while choking, marked daytime sleepiness, chest pain, fainting, difficulty breathing, or another acute concern should lead to medical or emergency help based on the symptoms. A camera or vision product marketed for infants cannot replace adult supervision, safe-sleep practice, or an authorized vital-sign monitoring plan, and video alone does not justify a claim to prevent sudden infant death.