Source-linked AI summary
Review of Fall Detection Techniques: A Data Availability Perspective
Shehroz S. Khan, Jesse Hoey
TL;DR
Falls are rare, creating limited or absent fall data for training and making standard supervised learning difficult to apply. The paper proposes a data-availability taxonomy, reviews methods within its categories, and identifies treating falls as abnormal activities as a plausible research direction.
Problem
Falls are rare events, so training data for them may be insufficient or unavailable, limiting the direct applicability of standard supervised machine learning methods.
Method
The paper proposes a fall-detection taxonomy based on fall-data availability during classifier training, independent of sensor type and specific feature extraction or selection methods.
Results
31.55 million normal activities per year are reported in comparison to only 2.6 falls, and the review identifies treating falls as abnormal activities as a plausible research direction.
Takeaways & Limitations
Considering a fall as an abnormal activity is identified as an important research direction for fall detection.
Takeaways & Limitations
Oversampling techniques may suffer from over-fitting.
Abstract
from arXiv · showhide
A fall is an abnormal activity that occurs rarely; however, missing to identify falls can have serious health and safety implications on an individual. Due to the rarity of occurrence of falls, there may be insufficient or no training data available for them. Therefore, standard supervised machine learning methods may not be directly applied to handle this problem. In this paper, we present a taxonomy for the study of fall detection from the perspective of availability of fall data. The proposed taxonomy is independent of the type of sensors used and specific feature extraction/selection methods. The taxonomy identifies different categories of classification methods for the study of fall detection based on the availability of their data during training the classifiers. Then, we present a comprehensive literature review within those categories and identify the approach of treating a fall as an abnormal activity to be a plausible research direction. We conclude our paper by discussing several open research problems in the field and pointers for future research.
1. Introduction
Falls are rare, diverse, and difficult to collect safely, leaving limited representative training data while missed detection creates health and safety risks. The paper therefore organizes fall-detection methods by fall-data availability, independently of sensors and feature-engineering choices.
- Missed abnormal activities, including falls, can impose health and safety risks and hinder independent living.
- Falls occur infrequently and diversely, creating insufficient training data and making them difficult to model with machine-learning and vision algorithms.
- Real-fall data collection can be harmful, unsafe, cumbersome, and subject to limited quantities and ethics clearances.
- Artificial falls can aid understanding but may not represent actual falls, and models trained on them may overfit and generalize poorly.
- 31.55 million normal activities per year contrast with only 2.6 falls under the paper’s nursing-home monitoring assumption.
- Standard supervised learning may be ill-posed for rare falls, motivating a taxonomy based on training-time fall-data availability.
- The taxonomy is independent of sensor type and feature extraction or selection, and supports a real-world perspective where falls are not abundant.
- The review surveys prior reviews, categorizes current methods using the taxonomy, examines abnormal-activity approaches, and identifies open research questions.
2. Survey of Existing Literature Review on Fall Detection
Existing reviews classify fall-detection methods mainly by sensors, features, algorithms, and design concerns, while repeatedly exposing the scarcity of real-world fall data. Reviews also report that simulated falls dominate the literature and that no standardized context-aware technique is widely accepted.
- Review taxonomies: Prior reviews organize fall-detection systems by wearable, ambient, vision, and kinematic sensors, alongside techniques such as motion, posture, proximity, inactivity, and body-shape analysis.
- System trade-offs: Reviews report that wearable and ambient devices are cheap and easy to install, whereas vision-based devices are more robust for detecting falls.
- Review scope: Reviews cover feature extraction and selection, classification algorithms, model evaluation, and practical concerns including privacy, occlusion, energy use, noise, and threshold selection.
- Data availability: Around 94% of surveyed studies use simulated falls, highlighting the difficulty of obtaining real fall data because falls are rare.
- Open issues: Despite many feature-extraction and machine-learning techniques, no standardized context-aware fall-detection technique is widely accepted by the research community.
- Open issues: Limited availability of real-life fall data is identified as a significant issue that could hinder system performance and complicates long-term experiments.
3. Taxonomy for the Study of Fall Detection
The proposed taxonomy organizes fall-detection classification methods by the amount of fall data available during training, independently of sensor type and feature engineering. It distinguishes direct fall detection when data are sufficient from methods that use limited data or model falls as abnormal activity when data are scarce.
- Taxonomy scope: The taxonomy is independent of the sensors used to capture human motion and the specific feature extraction or selection methods.
- Taxonomy structure: The taxonomy has two high-level categories: sufficient training data for falls and insufficient or no training data for falls.
- Sufficient fall data: With sufficient fall data, classifiers can use supervised machine learning, thresholds, or one-class methods, although cross-person generalization remains challenging.
- Insufficient fall data: With insufficient fall data, methods may use over/under-sampling, semi-supervised or cost-sensitive learning, anomaly detection, or one-class classification.
- Classification strategy: Category I attempts to detect falls directly, whereas Category II manipulates limited fall data or indirectly detects falls as abnormal activity.
- Abnormal-activity detection: These approaches require careful definitions of normal behavior because it varies across people and age groups, and not every deviation indicates a fall.
- Abnormal-activity detection: Abnormal-activity approaches learn normal behavior without fall samples, avoiding the fall-data imbalance but risking false alarms when normal behavior is poorly learned.
4. Literature Review
The review organizes fall-detection methods by the availability of fall data, covering supervised, sampling, semi-supervised, cost-sensitive, and one-class approaches. It highlights practical limitations including overfitting, imbalanced data, threshold selection, and poor generalization.
- Thresholding Techniques: Thresholding methods compare raw or processed sensor data with predefined thresholds, usually requiring fall data and domain knowledge or data analysis.
- One-Class Classification: One-class classifiers can model a fall region when sufficient fall data exists, but realistic deployment is constrained by the difficulty of collecting real falls.Reported methods include K-center, K-nearest neighbour, OSVM, and SCMPM; SCMPM achieved the best overall performance among four tested OCC methods.
- Sampling Techniques: Sampling techniques address class imbalance by oversampling falls or undersampling normal activities before training supervised classifiers.Oversampling can overfit when artificial points poorly represent actual falls, while excessive undersampling can underfit the normal class.
- Semi-Supervised Learning: Semi-supervised methods combine limited labelled data with abundant unlabelled activity data to reduce exhaustive manual labelling.Examples include decision-tree initialization followed by threshold-based learning, nearest-neighbour personalization, expert refinement, and automatic annotation with classifier retraining.
- Limitations: Artificial falls may cause overfitting, while traditional supervised classifiers can be dominated by the majority normal-activity class and yield unfavourable accuracies on imbalanced datasets.
- Cost-Sensitive Classification and Decision Theory: Cost-sensitive classification assigns different penalties to errors because false alarms and missed falls should not be treated equally.The review notes that most fall-detection studies do not incorporate the costs of reporting or failing to report a fall, and cost-ratio estimation can overfit without intuitive interpretation.
5. Detecting Falls in the Absence of their Training Data
When fall samples are insufficient or absent during training, the reviewed approaches detect falls indirectly as deviations from learned normal activity. The section covers one-class and anomaly-based methods across wearable, acoustic, and vision sensors, with mixed performance and thresholding challenges.
- Anomaly-Based Detection: Outlier-detection and one-class methods learn normal behaviour and flag deviations as abnormal activities rather than directly modelling fall samples.This strategy can be used when fall samples are unavailable during classifier training, but abnormal activities are not necessarily falls.
- Anomaly-Based Detection: Recent studies show that falls can be identified without prior fall examples or specific domain knowledge about fall patterns.
- Acoustic Methods: Acoustic systems trained on normal sounds reported different rankings: OSVM performed best in one study but was outperformed by supervised learning, while interference suppression reduced sensitivity to interference.
- Wearable Sensor Methods: For smartphone accelerometer data, OCNN outperformed OSVM but remained below supervised SVM, whereas one-class K-nearest neighbour was better than or equivalent to supervised SVM and KNN in most cases.
- Unseen-Fall Modelling: X-Factor HMMs use inflated covariance estimated from outliers in normal data as proxies for unseen falls, achieving high fall-detection rates on two datasets but increasing false alarms.
- Summary: Across anomaly-oriented fall detection and related abnormal-activity methods, OSVM, OCNN, HMM, GMM, and their variants are most common, alongside EVT, clustering, Dynamic Bayesian Networks, and SLDS.
6. Future Research Directions and Challenges
The paper identifies treating falls as abnormal activities as a plausible direction because falls are infrequent and real fall data is costly to collect. It also outlines challenges in costs, thresholds, personalization, evaluation, and cross-person generalization.
- Unseen Falls as Anomalies: Treating falls as abnormal activities reflects their infrequent occurrence and avoids relying exclusively on long-term collection of real fall data.Collecting real falls may require years, extensive resources, ethics clearances, and substantial analysis and labelling effort.
- Unseen Falls as Anomalies: The concept of normal activity must be clearly defined, and insufficient normal-activity data may produce excessive false alarms.
- Candidate Methods: Future work includes auto-encoders, recurrent neural networks, and Extreme Value Theory for learning representations, sequential patterns, or rare-event distributions without many fall examples.
- Decision Costs and Thresholds: A major unresolved issue is incorporating error costs and optimizing thresholds when the costs of false alarms and missed falls are poorly understood and fall data is unavailable.
- False-Alarm Reduction: Combining fall-risk scores with scores for unseen falls may help reduce false alarms, while static heuristic rules are hard to generalize.
- Evaluation and Generalization: Evaluation lacks a standard framework and shared repositories, and classifier performance decreases significantly when testing uses a different dataset from training.
Declaration
The paper declares no conflicts of interest and states that ethical approval was not required.
- The authors report no conflicts of interest.
- The authors state that ethical approval was not required.