Falls represent critical safety challenge, accounting for16% of non-fatal occupational injuries and 865 workplace fatalities in the United States in 2022. Traditional post impact fall detection systems prompt first aid after falls occur but fail to prevent injuries caused by ground impact. Pre-impact fall detection enables protective intervention before ground contact by activating safety systems, such as wearable airbags, to reduce the severity of impact. However, achieving this requires recognizing falls early enough to provide stuffiest time for deploying protective devices while maintaining detection accuracy and reliability on resource-constrained wearable platforms. This thesis addresses real-time pre-impact fall detection through three interconnected contributions advancing embedded wearable systems. First, a comprehensive dataset of 46.05 hours from 61 (+10) participants encompasses construction-specific elevation falls absent from existing benchmarks, combining controlled laboratory protocols with naturalistic construction site recordings. Integration with the KFall benchmark provides a robust foundation for model development and validation across diverse fall scenarios and operational contexts. Second, a lightweight convolutional neural network optimized for STM32F722 microcontrollers achieves 86.69% F1-score while operating within strict memory and processing constraints. The architecture addresses critical timing requirements ignored by existing methods, but severe class imbalance produces excessive false negatives of 4.17%, limiting practical reliability for safety-critical applications. Third, an enhanced two-stage detection pipeline addresses these limitations through generative data augmentation combined with hierarchical classification. Systematic evaluation of three generative models (Variational Autoencoder, Conditional TimeGAN (CT-GAN), and hybridVAE-GAN) identifies effective strategies for synthesizing realistic falling patterns while preserving temporal dynamics. Beyond segment-level classification, a novel event-level aggregation stage employs Random Forest classification on consecutive segment predictions. This supervised mechanism replaces simple threshold rules, learning confidence patterns that distinguish genuine falls from transient false positives. Validated through 5-fold subject-independent cross-validation, the complete pipeline achieves 99.5% event-level F1-score with 99.9% precision and false negatives below 1%. Post-training quantization enables deployment on embedded microcontrollers, meeting real-world timing constraints for wearable safety systems. Real-world validation using commercial airbag safety jackets demonstrates deployment readiness. This thesis bridges academic fall detection research and deployable embedded safety systems, advancing wearable sensing technologies for occupational safety, elderly care, and real-time human activity recognition under severe resource constraints.
Le cadute rappresentano una sfida critica per la sicurezza, essendo responsabili del 16% degli infortuni professionali non mortali e di 865 decessi sul lavoro negli Stati Uniti nel 2022. I tradizionali sistemi di rilevamento delle cadute post-impatto sollecitano l’intervento di primo soccorso dopo che la caduta si è verificata, ma non riescono a prevenire le lesioni causate dall’impatto con il suolo. Il rilevamento pre-impatto delle cadute consente un intervento protettivo prima del contatto con il suolo, attivando sistemi di sicurezza, come airbag indossabili, per ridurre la gravità dell’impatto. Tuttavia, ciò richiede il riconoscimento della caduta con sufficiente anticipo per consentire il dispiegamento dei dispositivi di protezione, mantenendo al contempo accuratezza e affidabilità del rilevamento su piattaforme indossabili con risorse limitate. Questa tesi affronta il rilevamento pre-impatto delle cadute in tempo reale attraverso tre contributi interconnessi che fanno progredire i sistemi indossabili embedded. In primo luogo, viene presentato un dataset completo di 46,05 ore raccolte da 61 (+10) partecipanti, che include cadute da dislivello specifiche del settore delle costruzioni, assenti nei benchmark esistenti, combinando protocolli controllati in laboratorio con registrazioni naturalistiche in cantieri edili. L’integrazione con il benchmark KFall fornisce una solida base per lo sviluppo e la validazione dei modelli in diversi scenari di caduta e contesti operativi. In secondo luogo, una rete neurale convoluzionale leggera, ottimizzata per microcontrollori STM32F722, raggiunge un F1-score dell’86,69% operando entro rigorosi vincoli di memoria ed elaborazione. L’architettura affronta requisiti temporali critici ignorati dai metodi esistenti; tuttavia, il forte sbilanciamento tra le classi produce un numero eccessivo di falsi negativi, pari al 4,17%, limitandone l’affidabilità pratica per applicazioni safety-critical. In terzo luogo, una pipeline di rilevamento avanzata a due stadi affronta tali limitazioni attraverso l’augmentazione generativa dei dati combinata con una classificazione gerarchica. La valutazione sistematica di tre modelli generativi — Variational Autoencoder, Conditional TimeGAN (CT-GAN) e ibrido VAE-GAN — identifica strategie efficaci per sintetizzare pattern realistici di caduta preservandone le dinamiche temporali. Oltre alla classificazione a livello di segmento, un nuovo stadio di aggregazione a livello di evento impiega una classificazione Random Forest sulle predizioni di segmenti consecutivi. Questo meccanismo supervisionato sostituisce semplici regole basate su soglie, apprendendo pattern di confidenza che distinguono le cadute reali dai falsi positivi transitori. Validata tramite cross-validation subject-independent a 5 fold, la pipeline completa raggiunge un F1-score a livello di evento del 99,5%, con una precisione del 99,9% e falsi negativi inferiori all’1%. La quantizzazione post-training consente il deployment su microcontrollori embedded, rispettando i vincoli temporali reali dei sistemi di sicurezza indossabili. La validazione in condizioni reali mediante giacche di sicurezza commerciali con airbag dimostra la prontezza del sistema per il deployment. Questa tesi colma il divario tra la ricerca accademica sul rilevamento delle cadute e i sistemi di sicurezza embedded effettivamente implementabili, contribuendo all’avanzamento delle tecnologie di sensing indossabile per la sicurezza occupazionale, l’assistenza agli anziani e il riconoscimento in tempo reale delle attività umane in condizioni di severe limitazioni di risorse.
Real-Time Pre-Impact Fall Detection in Wearable Systems: Lightweight Deep Learning Architectures for Embedded Deployment
Ali, Muhammad Toqeer
2026-01-01
Abstract
Falls represent critical safety challenge, accounting for16% of non-fatal occupational injuries and 865 workplace fatalities in the United States in 2022. Traditional post impact fall detection systems prompt first aid after falls occur but fail to prevent injuries caused by ground impact. Pre-impact fall detection enables protective intervention before ground contact by activating safety systems, such as wearable airbags, to reduce the severity of impact. However, achieving this requires recognizing falls early enough to provide stuffiest time for deploying protective devices while maintaining detection accuracy and reliability on resource-constrained wearable platforms. This thesis addresses real-time pre-impact fall detection through three interconnected contributions advancing embedded wearable systems. First, a comprehensive dataset of 46.05 hours from 61 (+10) participants encompasses construction-specific elevation falls absent from existing benchmarks, combining controlled laboratory protocols with naturalistic construction site recordings. Integration with the KFall benchmark provides a robust foundation for model development and validation across diverse fall scenarios and operational contexts. Second, a lightweight convolutional neural network optimized for STM32F722 microcontrollers achieves 86.69% F1-score while operating within strict memory and processing constraints. The architecture addresses critical timing requirements ignored by existing methods, but severe class imbalance produces excessive false negatives of 4.17%, limiting practical reliability for safety-critical applications. Third, an enhanced two-stage detection pipeline addresses these limitations through generative data augmentation combined with hierarchical classification. Systematic evaluation of three generative models (Variational Autoencoder, Conditional TimeGAN (CT-GAN), and hybridVAE-GAN) identifies effective strategies for synthesizing realistic falling patterns while preserving temporal dynamics. Beyond segment-level classification, a novel event-level aggregation stage employs Random Forest classification on consecutive segment predictions. This supervised mechanism replaces simple threshold rules, learning confidence patterns that distinguish genuine falls from transient false positives. Validated through 5-fold subject-independent cross-validation, the complete pipeline achieves 99.5% event-level F1-score with 99.9% precision and false negatives below 1%. Post-training quantization enables deployment on embedded microcontrollers, meeting real-world timing constraints for wearable safety systems. Real-world validation using commercial airbag safety jackets demonstrates deployment readiness. This thesis bridges academic fall detection research and deployable embedded safety systems, advancing wearable sensing technologies for occupational safety, elderly care, and real-time human activity recognition under severe resource constraints.| File | Dimensione | Formato | |
|---|---|---|---|
|
Muhammad_Toqeer_Ali_PhD_THESIS_2022_2025 copia_signed.pdf
embargo fino al 13/05/2027
Tipologia:
Tesi di dottorato
Licenza:
Dominio pubblico
Dimensione
2.2 MB
Formato
Adobe PDF
|
2.2 MB | Adobe PDF | Visualizza/Apri Richiedi una copia |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.



