Reliable localization of firefighters, especially in indoor and GPS-denied environments, remains a persistent challenge. Personal Dead Reckoning (PDR), powered by wearable inertial sensors, presents a viable solution but relies on resilient communication to transmit real-time position data to coordination centers. This paper examines the feasibility of using LoRa-based mesh networks as an infrastructure-free communication solution for wearable PDR systems in firefighting scenarios. We evaluate the performance of Meshtastic LoRa mesh nodes through controlled range tests and a live multi-floor residential firefighting exercise. Results indicate that LoRa mesh networks can meet the data needs of wearable PDR systems in most operational contexts, with minimal infrastructure and setup. We also propose practical deployment strategies, including firetruck-node integration and repeater placement, demonstrating that many building types can be covered with minimal effort. These findings support the development of robust, infrastructure-independent systems for on-body localization and mission-critical communication.
Wi-Fi signals can reveal patterns of human activity by analyzing how movement disturbs signal propagation. While prior work on Channel State Information (CSI) has shown promise for indoor activity recognition, it typically assumes sensor placement within the monitored environment or in adjacent rooms. This study investigates a more constrained and privacy-sensitive scenario: can activity inside a home be inferred using low-cost wireless devices placed entirely outside, with no interior access? Using a custom active sensing setup, we collected over 600 min of CSI data across two residential apartments, capturing room-level presence under varying wall materials, device placements, and participant behavior. We evaluated multiple learning strategies—including supervised, unsupervised, and semi-supervised models—to assess the feasibility of inferring interior activity from exterior signals. Results show that room-level localization is achievable even through thick residential walls, though performance varies substantially with environmental structure and sensor configuration. These findings demonstrate the viability of a previously unexplored form of passive activity inference and raise important questions about wireless privacy in domestic settings.
Population aging increases the need for accessible cardiac rhythm screening outside clinical settings. We present a proof-of-concept, low-complexity pipeline for at-home electrocardiogram (ECG) rhythm screening tailored to Ambient Assisted Living (AAL). To mimic single-lead devices, we extract interpretable features from lead II, including ventricular rate, QRS duration, QT/QTc, RR-interval statistics, and simple counts derived from wavelet-based delineation. We compare a compact Fully Connected Neural Network (FCN) against a tree-based model (XGBoost). Models are developed on ChapmanECG under an 11-class label space and evaluated in matched splits; XGBoost yields higher mean accuracy than the FCN. For an external, system-level check, we apply the trained pipeline to short excerpts from the MIT-BIH Arrhythmia Database and harmonize labels via a graded protocol (Correct/Partially Correct/Incorrect) to account for taxonomy differences. Under this protocol, the pipeline attains 82.98
This paper introduces a pipeline for generating synthetic Inertial Measurement Unit data to address data scarcity in Human Activity Recognition (HAR). The ability to differentiate between activities benefits fall detection, fitness tracking, and health analysis. However, limited availability and high cost of labeled training data hinders robust model development. While current synthetic data approaches remain cumbersome and require extensive domain expertise, this automated pipeline makes data generation more accessible, enabling researchers to develop more powerful HAR systems. The proposed methodological framework combines video generation models with pose estimation techniques and physics-based kinematic transformations to produce synthetic sensor data that effectively augments real-world measurements. Through evaluation across established benchmark datasets, results demonstrate that models trained on a combination of real and synthetic data achieve improved performance compared to baseline approaches, with F1-score improvements of up to 1.7
LiDAR-based point cloud analysis for Human Activity Recognition (HAR) is increasingly relevant in assistive technologies, particularly for supporting elderly and care-dependent individuals through applications such as fall detection. In this work, we present DenseVoxelNet3D, a compact and efficient 3D convolutional neural network designed for voxelized LiDAR input. The architecture combines a voxelization-based preprocessing pipeline with densely connected layers to improve feature propagation and mitigate vanishing gradient issues. By integrating global average pooling and dropout-based regularization, the model achieves strong generalization with minimal overfitting. On the HmPEAR dataset [14], DenseVoxelNet3D outperforms conventional architectures such as AlexNet and ResNet in terms of classification accuracy and computational efficiency. Our code is available at https://github.com/eckabeg/DenseVoxelNet3D.git .
