Occupancy Indicator: Sensing & Applications

Updated 31 May 2026

Occupancy indicators are signals representing the state of presence in a given space, utilizing sensor data and analytical methods.
They leverage modalities like video, environmental sensors, and RF to support applications in smart buildings, healthcare, and cyber-security.
Their design employs tailored algorithms and performance metrics to drive energy savings, resource management, and operational efficiency.

An occupancy indicator is a quantitative or qualitative signal designed to represent the state of “presence”—typically of people (or in some contexts, processes)—in a given spatial, cyber, or infrastructural domain. Occupancy indicators can be binary (“occupied” vs. “vacant”), categorical (multi-level presence), or continuous (fractional or absolute count), and are central to applications such as smart building management, energy optimization, healthcare resource allocation, and cyber-physical security. The technical formulation, sensor modality, and algorithmic mapping from raw signal to indicator are highly domain-dependent.

1. Canonical Definitions and State Spaces

An occupancy indicator formalizes the notion of presence over a defined space and time scale. In most building and control contexts, the occupancy indicator at time $t$ can be cast as:

A discrete scalar: $z_t \in \{0,1\}$ (unoccupied/occupied) or $z_t \in \{0,1,\dots,K\}$ (multi-class occupancy level),
A real-valued fractional ratio $r_t \in [0,1]$ (continuous proxy, e.g., fractional presence),
A vector $X_t \in \{0,1\}^{W \times H \times D}$ (spatial occupancy grid for each voxel, as in autonomous driving).

Formulations are tailored to sensing modality. For example, video-based occupancy detectors estimate the person count $\hat{y}_t$ and map this to a binary state $\hat{z}_t = \mathbb{I}[\hat{y}_t \geq 1]$ (Qaisar et al., 27 Mar 2026), while smart meter or CO $_2$ /VOC-based systems generate a binary label via an ML classifier thresholded at 0.5 probability (Luo et al., 2022, Varnosfaderani et al., 2022).

Table 1 summarizes indicator representations across major application domains:

Application	Indicator	Example Symbol
Building HVAC control	binary/count	$z_t,\,\hat{y}_t$
Health system stress (ICU)	integer-valued	$X(t)$
Automotive perception	3D binary grid	$z_t \in \{0,1\}$ 0
Cache timing channels	occupancy vector	$z_t \in \{0,1\}$ 1

2. Sensor Modalities and Measurement Pipelines

Occupancy indicators derive from diverse sensor streams and signal-processing pipelines:

Computer vision: Occupancy is inferred from person detection or tracking in RGB or thermal images. YOLOv5 on thermal images with transfer learning achieves $z_t \in \{0,1\}$ 2, $z_t \in \{0,1\}$ 3, $z_t \in \{0,1\}$ 4 for single-person office presence (Cui et al., 13 May 2025). Omnidirectional low-res video with embedded YOLOv2 provides room-level counts with $z_t \in \{0,1\}$ 5–0.88 (Callemein et al., 2020). LLM-based refinement and temporal consistency frameworks further enhance reliability and reduce missed presence events (Qaisar et al., 27 Mar 2026).
Environmental sensors: CO $z_t \in \{0,1\}$ 6, VOC, and light sensors, often fused via ML. Indicator is typically extracted using an SVM or Random Forest decision on features such as CO $z_t \in \{0,1\}$ 7 level, its first derivative, spatial concentration differences, and/or VOC concentration (Varnosfaderani et al., 2022, Huang et al., 2024).
Smart meter data: Aggregate power, sometimes with calendar/time features, subjected to deep CNN+attention processing or hybrid Transformer-LSTM pipelines. Occupancy indicator is a probability or binary flag per time window, e.g., ABODE-Net achieves $z_t \in \{0,1\}$ 8–0.86 on public datasets (Luo et al., 2022); hybrid Transformer-RNN approaches yield accuracy $z_t \in \{0,1\}$ 9 (Liang et al., 2023).
RF/BLE-based: Occupancy inferred from statistical patterns in Wi-Fi or BLE RSSI traces, with SVM, DT, and RF yielding $z_t \in \{0,1,\dots,K\}$ 0– $z_t \in \{0,1,\dots,K\}$ 1 accuracy for both binary indicator and small count estimation (MAE $z_t \in \{0,1,\dots,K\}$ 2) (Demrozi et al., 2021, Galluzzi et al., 2019).
Acoustic/reverberation: Features such as RT $z_t \in \{0,1,\dots,K\}$ 3 used as distinguishing features within inductive decision trees, achieving $z_t \in \{0,1,\dots,K\}$ 4 accuracy in classroom detection (Jain et al., 2016).
Healthcare capacity: Occupancy indicator is the number of ICU beds occupied, modeled as a continuous-time Markov process (immigration–death), directly informing epidemic control (Awasthi et al., 2022).
Cybersecurity (cache): Indicator is per-core cache occupancy, sampled at kHz, with zero-lag cross-correlation used to detect covert channels (Yao et al., 2019).

3. Algorithmic and Statistical Mapping

The conversion from raw sensor traces $z_t \in \{0,1,\dots,K\}$ 5 to occupancy indicator $z_t \in \{0,1,\dots,K\}$ 6 or $z_t \in \{0,1,\dots,K\}$ 7 is highly context-specific. Approaches include:

Direct ML classification: $z_t \in \{0,1,\dots,K\}$ 8, with $z_t \in \{0,1,\dots,K\}$ 9 a hand-crafted or learned feature transformation. Classifiers include SVM (RBF), decision trees, random forests, logistic regression, and deep learning architectures (CNN, RNN, transformer hybrids) (Luo et al., 2022, Liang et al., 2023, Varnosfaderani et al., 2022).
Dynamic models/regression: For count estimation, especially from CO $r_t \in [0,1]$ 0, mapping is often via learning $r_t \in [0,1]$ 1 such that $r_t \in [0,1]$ 2, where $r_t \in [0,1]$ 3 is a concatenation of recent CO $r_t \in [0,1]$ 4, prior occupancy, and venting levels. Feature Scaled ELM, seasonal-trend decomposition plus regression, and hybrid deep networks are used for robust head-counting, with postprocessing corrections for “empty” intervals (Jiang et al., 2016, Arief-Ang et al., 2017).
Stochastic and EM-inspired frameworks: In occupancy extraction from aggregate power, probabilistic profile generators coupled to interpretable load disaggregators (e.g., Kolmogorov–Arnold Networks) jointly infer discrete and continuous occupancy, maximizing posterior likelihood over observed loads (Zhang et al., 23 Apr 2025).
Thresholding and rules: In low-complexity systems, indicators are threshold comparisons (e.g., $r_t \in [0,1]$ 5), or multi-sensor rules with statically chosen splits (Jain et al., 2016, Huang et al., 2024).
Cross-correlation for mutual gain-loss detection: In cache occupancy, the indicator is evaluated not just per-process but by the cross-correlation $r_t \in [0,1]$ 6 between changes in two processes' cache block ownership (Yao et al., 2019).

4. Quantitative Performance and Validation Metrics

The reliability of an occupancy indicator is assessed by metrics tailored to task and granularity:

Binary classification: Accuracy, precision, recall, $r_t \in [0,1]$ 7, ROC AUC on ground-truth presence (Luo et al., 2022, Cui et al., 13 May 2025, Qaisar et al., 27 Mar 2026).
Counting: MAE, RMSE, and $r_t \in [0,1]$ 8-tolerance accuracy (fraction of predictions within $r_t \in [0,1]$ 9 occupants) (Jiang et al., 2016, Huang et al., 2024).
Time-series consistency: Temporal smoothing, ID-switch and fragmentation rates (video/tracking), delay in response to transitions (Callemein et al., 2020, Qaisar et al., 27 Mar 2026).
Operational impact: Energy savings associated with occupancy-driven control (e.g., $X_t \in \{0,1\}^{W \times H \times D}$ 0% reduction in HVAC energy (Qaisar et al., 27 Mar 2026); literature cites up to 44% (Cui et al., 13 May 2025)), improvements in comfort, or resource-planning (average ICU stay, forecast intervals) (Awasthi et al., 2022).
Adversarial robustness: Resistance to spoofing or adaptive attacks, e.g., occupancy-based cache timing channel detectors remain robust at $X_t \in \{0,1\}^{W \times H \times D}$ 1 even with sophisticated noise injection, whereas miss-rate detectors collapse ( $X_t \in \{0,1\}^{W \times H \times D}$ 2) (Yao et al., 2019).

5. Privacy, Computational Efficiency, and Deployment

Occupancy indicators entail trade-offs in privacy, computational load, and ease of deployment:

Enhanced privacy: Low-resolution (128×96) thermal or omnidirectional video ensures facial identity is irrecoverable (Cui et al., 13 May 2025, Callemein et al., 2020). RGB-based detectors raise privacy concerns and are avoided in privacy-sensitive deployments.
Embedded/Edge feasibility: Compressed network backbones (YOLOv5s, YOLOv2) with input resolution $X_t \in \{0,1\}^{W \times H \times D}$ 3 MP, small model size ( $X_t \in \{0,1\}^{W \times H \times D}$ 47 MB), and $X_t \in \{0,1\}^{W \times H \times D}$ 5 fps CPU inference enable real-time edge use for control integration (Cui et al., 13 May 2025, Callemein et al., 2020).
Sensor cost and ubiquity: Wi-Fi/BLE-based indicators leverage existing hardware, providing room-level occupancy at sub-\$X_t \in \{0,1\}^{W \times H \times D} $6_2$ sensors suffice for most small, well-mixed rooms; multi-sensor arrays or additional VOC/light signals offer marginal improvements (Huang et al., 2024, Varnosfaderani et al., 2022).
Scalability: Model architectures relying only on aggregate data (CO $X_t \in \{0,1\}^{W \times H \times D}$ 7, smart meter) can be deployed at scale with retraining or transfer learning (Luo et al., 2022, Zhang et al., 23 Apr 2025).
Privacy–accuracy balance: BLE/Wi-Fi systems are non-intrusive and GDPR-compliant but may under-/overcount due to device-user mismatch; image-based systems are more accurate but must mitigate privacy via resolution as above.

6. Applications Across Domains

Occupancy indicators are foundational in:

Smart building automation: Dynamic HVAC/setpoint control, lighting, security, demand response via continuous or binary occupancy signals (Cui et al., 13 May 2025, Qaisar et al., 27 Mar 2026).
Healthcare resource monitoring: ICU occupancy as a real-time system stress indicator, supporting capacity planning, public policy, and outbreak control using stochastic immigration–death models (Awasthi et al., 2022).
Cybersecurity: Cache occupancy indicators for the detection of covert timing channels and information leakage—leveraging mutual eviction patterns rather than spoofable cache misses (Yao et al., 2019).
Automotive perception: Occupancy grids in voxel space for forecasting and multi-agent coordination; hybrid metrics beyond human annotation, evaluating spatial and temporal plausibility (Wang et al., 31 Mar 2025).
Smart mobility: Binary occupancy indicators at EV charging stations inform MDP-based routing and queuing for optimized energy and journey planning (Dastpak et al., 2023).

7. Limitations, Robustness, and Research Directions

Key limitations include:

Modality sensitivity: CO $X_t \in \{0,1\}^{W \times H \times D}$ 8-based indicators degrade with poor air mixing or open windows; video-based occupancy is constrained by occlusion, orientation, and lighting (Cui et al., 13 May 2025, Huang et al., 2024).
Transferability: Occupancy indicators trained for single-occupant offices may not generalize to multi-occupancy or open-plan settings without additional model adaptation (Cui et al., 13 May 2025, Qaisar et al., 27 Mar 2026).
Temporal lags: Passive environmental indicators (CO $X_t \in \{0,1\}^{W \times H \times D}$ 9, VOC) can lag real occupancy transitions by several minutes, impacting real-time control.
Adversarial resistance: Indicators that directly reflect resource use or mutual interference (e.g., cache occupancy) are inherently robust to evasion, whereas others (miss-rate, signal thresholding) are more easily subverted (Yao et al., 2019).
Future work: Multimodal fusion, semi-supervised training, domain adaptation, and integration with temporal reasoning continue to be active areas. Expanding to occupant counting, activity recognition, and scalable, privacy-preserving learning are also in focus.