Definition, input, and task boundaries
Wi-Fi channel state information sleep body-movement recognition algorithms use changes in radio propagation among a transmitter, receiver, and indoor reflection paths to determine whether a person in bed has moved and may further classify minor movement, turning, sitting up, bed exit, or an unknown event. Human movement changes the amplitude and phase of direct and multipath components, providing motion cues without a camera or a body-worn sensor.
CSI commonly provides a complex channel response indexed by antenna pair, subcarrier, packet, and time, making it more detailed than RSSI, which compresses the channel into a single strength value. It is not a direct image of body contours, joint trajectories, or sleep stages. An algorithm should identify the wireless protocol, band, bandwidth, transmitting and receiving devices, antenna configuration, packet source, and available CSI dimensions rather than treat outputs from different hardware as identical inputs.
“Body movement” needs an operational definition. It may mean any nonrespiratory movement, activity above a duration and magnitude threshold, or an event with a start, end, and class. A radio disturbance around a bed does not prove that a particular sleeper turned, and it does not diagnose periodic limb movement, a seizure, or an arousal. Clinical and safety uses require synchronized reference signals and independent validation of the final task.
Acquisition link, amplitude and phase, and calibration
CSI is affected by transmit power, automatic gain, carrier-frequency offset, sampling-frequency offset, packet-detection delay, and phase wrapping. Raw amplitude can jump when gain changes, while raw phase may contain linear trends and random offsets unrelated to the person. Preprocessing should record and correct identifiable hardware errors while retaining the correction method, failure state, and native timestamps; it must not attribute every change to body movement.
Packets commonly arrive at irregular intervals and may be lost because of network load, interference, or power-saving behavior. Interpolation and resampling can create a regular time axis but alter event edges and spectra, while active probe traffic occupies airtime and remains subject to channel-access rules. The algorithm should report the distribution of actual packet rate, loss, timestamp accuracy, and effective throughput and verify the required temporal resolution under realistic network traffic.
Locations of the transmitter, receiver, and bed determine the sensing region and multipath geometry. Moving a router, rotating an antenna, relocating the bed, opening a door, or adding furniture can reshape the baseline, while neighbors and other radio devices can perturb the link. Installation should preserve room and link configuration, perform empty-room and occupied-rest calibration, and trigger a quality warning or revalidation after a material environmental change.
Preprocessing, link selection, and event segmentation
Preprocessing may include outlier-packet removal, phase unwrapping, amplitude normalization, detrending, low- or band-pass filtering, principal components, antenna ratios, and time-frequency transforms. Each step can amplify noise or remove a slow turn, so parameters should match the target motion scale. If normalization uses whole-night statistics, future observations must not leak into real-time output or test windows.
Multiple antennas and subcarriers provide many correlated sequences, but not every link crosses the bed region or remains sensitive in the current posture. Link selection may use stability, movement response, signal-to-noise, or learned weights and must not choose the best channels after examining the test labels. Fusion should retain antenna-pair and subcarrier provenance and test whether a single failure, deep fade, or radio interferer dominates the result.
Event segmentation often uses short-window variance, differential energy, correlation change, signal subspaces, or learned models to locate onset and offset. Window length, overlap, hysteresis, minimum duration, and merging of adjacent events directly change event count and duration. Breathing, bed vibration, and multipath recovery after movement can produce tails, so an algorithm should allow transition or unknown states instead of splitting one turn into several events.
Representation, classification, and temporal context
Movement representations may use amplitude and phase statistics, subcarrier correlation matrices, Doppler or time-frequency energy, duration, and spatial link change, or they may be learned by convolutional, recurrent, attention, or self-supervised models. Large movements are readily distinguished by energy, whereas a hand or foot movement, bedding motion, and a small posture adjustment overlap more. The model’s resolution should match label granularity; coarse training labels cannot support precise action names.
Sleep movements have temporal context. A turn often includes preparation, main movement, and stabilization; sitting up may precede bed exit; and periodic movement requires a sequence of event intervals. A state machine or sequence model can reduce frame-to-frame flicker, but priors must not invent actions that were not observed. Output should preserve event probability, class confidence, boundary uncertainty, and a model-rejection result.
Weak periodic change from breathing can serve as a resting background and can also be masked around light movement. Movement recognition may share quality information with respiration estimation, but disappearance of a respiratory pattern is not by itself movement, and broadband movement change is not a respiratory abnormality. If the output later supports sleep staging, the complete staging model needs validation; movement accuracy does not establish sleep-stage accuracy.
Labels, multi-person settings, and confounders
Training labels may come from synchronized infrared video, a pressure mat, an accelerometer, manual polysomnographic annotation, or multisource consensus. Each reference observes something different: wrist acceleration does not fully represent a torso turn, video can be obscured by bedding, and human annotators may disagree on event boundaries. The labeling protocol should define body region, classes, onset and offset, minimum event, and unknown status and should measure inter-rater agreement.
A person walking in the room, a bed partner, pet, fan, curtain, robot, or adjustable bed can all alter CSI. A single link usually cannot identify who moved, and multilink localization or source separation can still fail with occlusion, crossing paths, and simultaneous movement. The system should assess occupancy and assignability before emitting person-specific events and should retain multi-person mixture as a separate state.
Bedding thickness, sleep posture, body size, bed material, device height, and non-line-of-sight paths change the signature of the same movement. Network channel, firmware, and automatic frequency selection may also change overnight. Training should represent these factors and retain metadata. If the algorithm adapts rapidly in a new room, adaptation data, update limits, and negative-transfer detection must remain separate from final testing.
Validation, generalization, and reporting
Data should be partitioned by participant, room, and acquisition period so adjacent windows from one night cannot cross between training and test sets. Random window splits let a model memorize room baseline, person-specific patterns, and radio noise. Evaluation should include unseen people, unseen rooms, hardware changes, and different days or seasons and should distinguish zero-shot deployment, limited calibration, and full retraining.
Event detection should report sensitivity, positive predictive value, matching tolerance, timing error, false alarms per hour, and valid coverage. Multiclass recognition also needs a confusion matrix, macro-averaged measures, and unknown rejection. Frame accuracy is dominated by long periods of rest and cannot represent movement performance alone. Results should be stratified by movement size, posture, distance, line-of-sight condition, single or multiple occupants, and room.
Stress tests should cover packet loss, traffic bursts, channel switching, device restart, furniture change, a passerby, and simultaneous action. Algorithm comparisons need fixed data versions, labels, acquisition protocols, and scoring units, with computational latency and resource use reported. External evaluation should use independent sites and hardware and count failed nights, unavailable output, and recalibration frequency instead of retaining only successful segments.
Deployment, privacy, and safety boundaries
On-device processing can upload only events and quality information and thereby reduce raw CSI exposure, but radio sensing may still reveal occupancy, routines, and activity. The system should provide clear notice, appropriate consent, identity and access control, data minimization, encryption, retention, and deletion. Cohabitants and visitors can also be sensed, so “camera-free” must not be equated with absence of privacy risk.
Deployment should lock or record band, bandwidth, antennas, firmware, and link roles while monitoring packet rate, baseline drift, and environmental change. Before and after model updates or automatic adaptation, a fixed benchmark should test false alarms, misses, coverage, and subgroup performance, with rollback available. Prolonged low quality, device movement, or inseparable occupants should make the interface unavailable rather than preserve a stale state.
Wi-Fi CSI movement recognition can provide contactless activity trends and research features, but it cannot independently establish sleep, arousal, periodic limb movement, or another disease event. Unusually frequent movement accompanied by injury risk, abnormal breathing, or prominent daytime symptoms warrants appropriate further assessment. A trustworthy system separates radio change, inferred movement, and clinical interpretation and permits human review of synchronized evidence.