PoHAR: Hyperlocal Activity Recognition with Pollution Sensors

Key Features

---

Uniqueness: This study uniquely repurposes commodity low-cost air quality sensors, originally meant for pollutant tracking, as a privacy-preserving substrate for hyperlocal Human Activity Recognition (HAR). Unlike camera, microphone, RF, or wearable-based systems, PoHAR runs entirely on resource-constrained ESP32 microcontrollers and lets a distributed sensor network autonomously elect leaders, share self-supervised embeddings via a conflict-free replicated data type, and cluster only the sensors actually affected by an activity. This selective, on-device, leader-based approach allows multiple simultaneous indoor activities to be detected accurately in different zones of a home without centralized aggregation or high-fidelity data transmission.

• Conflict-free Replicated Set (Set-CvRDT): Implements a distributed set data structure with add-set, rem-set, and main-set primitives that guarantees consistent, conflict-free sharing of SSL embeddings across sensors over lossy UDP, even under node failure and message reordering.
• Self-Supervised Pollution Embeddings: Uses a time–frequency consistency (TF-C) based SSL encoder to convert raw, unlabeled multivariate air-quality time-series into low-dimensional embeddings, removing the need for large labeled activity datasets.
• Pollution-aware Hierarchical Clustering: Runs a partition-based distributed agglomerative clustering algorithm directly on ESP32 nodes, using distance bounds and mutual nearest-neighbor checks to dynamically identify only the sensor groups affected by a given activity.
• RAFT-based Leader Election: Employs a lightweight, fault-tolerant RAFT consensus mechanism for both network-wide coordination and per-cluster inference leadership, ensuring a single stable leader despite message loss or intermittent connectivity.
• On-device Hyperlocal Inference: Deploys off-the-shelf ML classifiers (Decision Tree, Random Forest, Extra Trees, Gaussian Naive Bayes, lightweight Neural Networks) directly on leader ESP32 nodes, achieving 97.41% accuracy for indoor activity and 99.68% for cooking activity recognition with inference latency below 34 μs.
• Large-Scale Real-World Validation: Evaluated across 30 diverse indoor sites over six months, collecting 89.1 million samples (13,646 hours) and 3,957 annotated activity events spanning occupancy, behavior, and eleven distinct cooking food items.

Potential Applications: The framework can be applied in various contexts, including:

• Smart Home Activity Monitoring: Enables privacy-preserving detection of occupant behaviors (entering/exiting, fan/AC usage, gathering, eating) without cameras or microphones.
• Remote Elderly and Healthcare Monitoring: Supports unobtrusive activity and routine tracking in assisted living or healthcare settings where visual surveillance raises privacy concerns.
• Kitchen and Cooking Safety Systems: Detects cooking type and specific food items being prepared, useful for dietary tracking, fire-risk alerts, or smart kitchen automation.
• Resource-Constrained IoT Deployments: Provides a blueprint for distributed consensus, clustering, and inference on low-power microcontroller networks beyond air quality, applicable to other dense sensor-network scenarios.

Contributors

Teaser Video

Publications

---

1. Karmakar, P., Reddy, K., and Chakraborty, S., PoHAR: Understanding Hyperlocal Human Activities with Pollution Sensor Networks. The 22nd Annual International Conference on Distributed Computing in Smart Systems and the Internet of Things (DCOSS-IoT) 2026