Aug 2026· Italian National Conference on Sensors· Vol 26· 0 citations· 43 references
Medicine
TL;DR
The findings show that CNN-based architectures dominate algorithm choice, edge devices dominate deployment platforms, and optimization remains central to real-time inference on constrained hardware.
Abstract
Falls represent a threat to older adults, overload healthcare systems, and reduce quality of life. Vision-based fall detection has advanced recently through deep learning, yet most proposed models lack validation on physical hardware and do not report inference-time metrics. This systematic review, following PRISMA and Kitchenham guidelines, targets this gap. We focus exclusively on vision-based systems that report inference speed on a specified device. We define real-time performance using a threshold of 10 fps, based on the reported duration of the critical fall phase in real-life falls. From 588 records across IEEE Xplore, ACM Digital Library, Web of Science Core Collection, and PubMed (2019–2024), only 11 met all inclusion criteria, highlighting how few studies validate real-time performance on physical hardware. The findings show that CNN-based architectures dominate algorithm choice, edge devices dominate deployment platforms, and optimization remains central to real-time inference on constrained hardware. Across these studies, we identify two persistent limitations: no real-world testing with older adults and reliance on small, controlled datasets with simulated falls.
Falling detection is vital for elderly care and intelligent surveillance; however, prevailing vision-based approaches predominantly frame it as static pose classification or discrete temporal pattern matching, fundamentally overlooking the instability dynamics of the human support system. This paper proposes a physics-informed falling detection framework that recasts falling as a stability-loss event in a coupled dynamical system. We introduce a novel dual-LTC architecture comprising a Center-of-Mass (CoM) subsystem and a Base-of-Support (BoS) subsystem, both instantiated as Liquid Time-Constant (LTC) neural networks to continuously model inertial trajectory evolution and ground-contact adjustment through adaptive time constants, Physical interpretability of falling motion. A learnable coupling module emulates physical interaction between the two subsystems, while a Stability Manifold classifier operates in the joint latent space to detect boundary crossing via Lyapunov-inspired stability metrics. Complementary counterfactual trajectory projection and Time-to-Collision (TTC) estimation further enable irreversibility assessment and early warning. The architecture is designed to support a three-state prediction paradigm (Normal, Falling, Fallen); in this preliminary study, we validate the core stability discrimination capability on a two-class dataset (Normal vs. Falling), leaving the full three-state temporal transition to future work. Unlike conventional CNN--RNN pipelines, the proposed formulation encodes continuous-time mechanical inertia, yielding a sub-50K-parameter network capable of real-time inference on resource-constrained edge devices. Extensive experiments demonstrate competitive accuracy with superior physical interpretability, validating its efficacy for low-compute visual fall detection.
Wenjun Xia, Zhicheng Peng, Haopeng Li et al.· 0 citations
Falls among older adults can cause serious injury and loss of independence. cStick 2.0 is a vision-enabled, IoMT-edge based smart walking-stick prototype that combines multimodal fall-risk classification with embedded obstacle awareness. The fall-risk classifiers were evaluated using a 9670-record development dataset, on which the compact DNN achieved 95.40% accuracy, 92.35% balanced accuracy, a macro F1-score of 93.90%, and a ROC-AUC of 97.44%. The Arduino Nicla Vision obstacle module used an INT8 Edge Impulse model with centroid-based direction assignment and time-of-flight distance sensing; 144 controlled trials produced 75.00% obstacle-presence accuracy at approximately 19–20 FPS. Sensor acquisition, GPS, display output, buzzer response, and CSV record accumulation were demonstrated at a prototype level. Synchronized older-adult evaluation, device-to-application communication, secure caregiver services, multimodal accessibility feedback, and longitudinal personalization remain future validation stages.
Laavanya Rachakonda, Sai Sri Harsha Chakravarthula, Saraju P. Mohanty et al.· Electronics· 0 citations
Alzheimer's disease affects millions of older adults worldwide, with wandering and fall-related injuries ranking among the most preventable yet dangerous consequences of cognitive decline. Current assistive technologies are largely reactive, alerting caregivers only after an incident occurs rather than anticipating the behavioral patterns that precede it. This paper presents a wearable IoT system that addresses this gap through continuous nine-axis inertial sensing, low-power Lo-RaWAN transmission, and a cloud-hosted inference pipeline for real-time fall detection. We evaluate four offline deep learning architectures (LSTM, GRU, 1D-CNN, and Transformer) alongside three online adaptive classifiers on the KFall benchmark dataset comprising 189,987 sliding windows with a 7.56% fall ratio. The stateless LSTM achieves the highest F1-score of 0.8893 with zero missed falls, while the 1D-CNN records the best precision at 0.9507 and the lowest false-alarm count. Beyond detection, the system incorporates a Retrieval-Augmented Generation (RAG) module backed by a Pinecone vector store and BART-Large generation model, enabling caregivers to query patient history and clinical guidance through natural language. All sensor payloads are protected with AES-256 encryption prior to transmission. The results show that pairing non-invasive wearable sensing with sequence models and a conversational caregiver interface produces a practical platform for keeping Alzheimer's patients safer.
Naman Omar, Vengadeswaran S, M. M· International Conference on...· 0 citations
This paper proposes a data-driven FFH detection method that integrates multiple complementary features into a unified score-based model, achieving a practical balance between detection sensitivity, false-positive suppression, computational efficiency, and real-time feasibility.
SangHyeok Kim, Daejin Park, Soon Ju Kang· Big Data and Cognitive Compu...· 0 citations
Real-time mask-wearing compliance monitoring with autonomous surveillance plays a key role in decision-making in industrial, healthcare, and educational systems. Existing face mask-wearing monitoring systems often rely on computationally intensive or cloud-based models, making them unsuitable for accurate, low-latency, real-time deployment on resource-constrained mobile and edge devices, particularly for multiclass mask-wearing compliance detection. In this regard, the primary focus is on an efficient autonomous edge vision model with a lightweight architecture. The proposed lightweight edge vision framework addresses the limitations of existing cloud-dependent and computationally intensive face mask-wearing monitoring systems by enabling accurate, low-latency, real-time multiclass compliance detection on resource-constrained mobile and edge devices. It consists of real-time data pipeline design, MobileNetV2-based model development, hyperparameter optimization, and real-time performance evaluation. It is implemented with a lightweight CNN architecture optimized for resource-constrained environments, which classifies into three categories: correctly masked, unmasked, and improperly masked. This is demonstrated by deploying the model on real-time devices, such as mobile devices and camera modules, that capture video streams, thereby addressing challenges such as varied lighting conditions and facial orientations. The proposed system achieves a high accuracy of 99.35% while maintaining low latency, making it suitable for public health surveillance in crowded settings. The findings highlight the potential of edge-based AI in enhancing compliance with safety protocols in public spaces.
Myneni Madhu Bala, Sreelakshmi Doma, Mohammad Riyaz Belgaum et al.· International journal of com...· 0 citations
Falls are a major health concern for older adults, and wearable sensors have been widely explored for detecting falls and enabling timely intervention. However, real-world falls are extremely rare: collecting 100 of them requires an estimated 100,000 days of monitoring, resulting in severely limited labelled data for training machine learning models. Consequently, many approaches rely on simulated datasets, often reporting high laboratory performance but limited real-world generalisation. We present a systematic evaluation of motion representations for wearable fall detection under real-world data scarcity. Using accelerometer signals, we compare interval-based, kernel-based, symbolic, and foundation model representations. As an interpretable baseline, we additionally investigate a lightweight symbolic representation that converts short motion segments into symbolic sentences augmented with physically-grounded impact descriptors. Experiments use FallAllD, a simulated falls dataset, and FARSEEING, a clinically verified real-world falls dataset. Through cross-validation, controlled data scarcity, and cross-dataset transfer, we examine how representation choices affect robustness under realistic deployment. Our results reveal that highly parameterised kernel and foundation models excel on simulated data but degrade severely under both data scarcity and domain shift. Although the interval-based representation achieves the strongest absolute real-world performance, augmenting a symbolic representation with physically-grounded impact descriptors yields the smallest degradation under domain shift and retains detection sensitivity under extreme scarcity, albeit at lower precision. These findings highlight the importance of evaluating beyond simulated benchmarks and show that representation choice is critical for deployable fall detection given the scarcity of real-world data.
Timilehin B. Aderinola, I. D’Ascanio, L. Palmerini et al.· 0 citations
