The dissertation is devoted to the development and study of machine learning methods based on the Transformer model for human fall detection using inertial measurement unit (IMU) sensor data and video. Falls are a leading cause of injury, hospitalisation and mortality among older adults, and a significant concern in occupational safety and smart environments. Existing fall-detection systems built on threshold algorithms and on convolutional and recurrent neural networks achieve high accuracy on individual datasets but suffer from high false-alarm rates, poor generalisation to previously unseen subjects, a need for large volumes of labelled data, privacy concerns in video-based approaches, and high computational cost that hinders real-time deployment on resource-constrained devices. The Transformer architecture, through its self-attention mechanism, models the long-range dependencies that are important for detecting short but contextually complex events such as falls.
The aim of the work is to improve the robustness, privacy and generalisation of fall-detection systems by developing a family of Transformer-based models, a balanced dataset, and real-time deployment strategies. The object of study is the process of automated human fall detection from inertial sensor data and video; the subject is machine learning methods and models based on the Transformer architecture for fall detection, in particular subject-independent and unsupervised detection, accounting for privacy and hardware constraints. The work employs methods of machine and deep learning, digital signal and time-series processing, computer vision, probability theory and mathematical statistics, extreme value theory (EVT), and standard evaluation protocols (Leave-One-Subject-Out, ROC/PR analysis).
A balanced multimodal dataset of 8,953 labelled samples (2,791 falls and 6,162 activities of daily living) from 29 participants was collected, and it is shown that adding barometric data to accelerometer data improves classification accuracy. A privacy-preserving video-based system based on EfficientNet for spatial feature extraction and LSTM for temporal modelling with a dynamic blurring module is proposed, anonymising the person without a substantial loss of detection accuracy. A comparative benchmark of seven Transformer architectures (Informer, Performer, Linformer, Temporal Convolutional Transformer and others) under unified conditions is provided, where the best models reach recall close to 100%. A Transformer model is deployed in real time on a System-on-Chip (Raspberry Pi Zero 2 W) with about 34 ms inference latency and an integrated cybersecurity layer (AES-256 encryption, blockchain). A Vision Transformer with spatio-temporal self-attention is applied for robust video analysis, while attention maps provide interpretability. A subject-independent unsupervised fall-detection method based on a masked Transformer and a combined energy-reconstruction scoring function with extreme-value-theory thresholding is developed; trained solely on activities of daily living, it reaches a ROC-AUC of 0.985, a PR-AUC of 0.955 and 95.8% event-level sensitivity under a leave-one-subject-out protocol.
The scientific novelty is that, for the first time, a subject-independent unsupervised fall-detection method based on a masked Transformer and a combined energy-reconstruction scoring function with extreme-value-theory thresholding has been developed, which requires no labelled fall examples; and, for the first time, a comparative benchmarking methodology for Transformer architectures for accelerometer-based fall detection under unified experimental conditions has been proposed. An architectural solution for real-time fall detection on a System-on-Chip with cybersecurity mechanisms, and a Vision Transformer based approach to video fall detection, have been improved. Privacy-preserving video analysis methods based on EfficientNet and LSTM, and approaches to building balanced multimodal datasets, have been further developed.
The practical significance is that the proposed methods can be applied in elderly monitoring systems, healthcare, occupational safety, smart homes, wearable devices and privacy-preserving video surveillance. The collected dataset and open source code ensure reproducibility of the results. The practical value is confirmed by a Ukrainian utility model patent for a device for fall detection and automatic notification of relatives.
Keywords: fall detection, Transformer model, machine learning, deep learning, time series, IMU, Vision Transformer, computer vision, unsupervised learning, anomaly detection, subject-independent detection, privacy preservation, edge computing, embedded systems.
