Review of Fall Detection Techniques: A Data Availability Perspective (1605.09351v2)

Abstract: 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.

• The paper presents a taxonomy classifying fall detection techniques based on fall data availability during the training phase.
• It emphasizes that treating falls as anomalies is a promising approach for addressing the data scarcity challenges of fall detection.
• This perspective advocates for developing data-efficient methods like One-Class Classification and anomaly detection for practical systems.

Fall Detection Techniques: A Data Availability Perspective

The paper "Review of Fall Detection Techniques: A Data Availability Perspective" provides a comprehensive taxonomy for fall detection strategies, emphasizing the availability of fall data during the training phase. This approach advocates for a nuanced understanding of fall detection mechanisms, distinct from conventional methods assuming data abundance.

The authors initiate the discourse by clarifying the significance of fall detection in human activity recognition. Falls, an occasional but critical event, pose substantial health risks, especially among the elderly. The sporadic nature of falls leads to considerable challenges in data acquisition, thereby complicating the application of standard supervised learning approaches. This necessitates the exploration of innovative methodologies that account for the data imbalance between falls and regular activities.

A pivotal contribution of this work is its proposed taxonomy, which classifies fall detection techniques based on the availability of fall data. The taxonomy is unrestricted by sensor type or specific feature extraction methodologies, broadening its applicability. Two primary categories emerge: scenarios with sufficient fall data and those with insufficient or absent fall data.

1. Sufficient Fall Data: This category encompasses supervised machine learning, threshold-based methods, and one-class classification (OCC) techniques trained on fall data. These approaches rely on adequate fall samples, often from artificial simulations, to inform classifier training. However, such simulated data may not accurately reflect real-world falls, leading to issues with overfitting and diminished generalizability across populations.
2. Insufficient or No Fall Data: This category includes a range of innovative techniques:
   • Over/Under-sampling: Adjusts class distributions to mitigate imbalance.
   • Semi-supervised Learning: Leverages a mix of labeled and unlabeled data to improve learning efficiency.
   • Cost-sensitive Learning: Implements differentiated misclassification costs to handle class imbalance.
   • Anomaly and Outlier Detection: Identifies falls as deviations from a learned model of normal activities.
   • OCC Techniques: Trains models using only normal activity data, identifying falls as anomalies.

The paper highlights a research trajectory where falls are treated as anomalies, reflecting a plausible shift from fall-as-event to fall-as-phenomenon. This perspective simplifies the architectural requirements to capture what is essentially a divergence from known behavioral norms. However, defining normal behavior robustly remains a crucial challenge.

The implications of this work are manifold, affecting both theoretical exploration and practical implementation. In terms of theory, it urges a re-thinking of the learning mechanisms in imbalanced settings, pushing the boundaries of OCC and anomaly detection platforms. Practically, it influences how fall detection systems are designed, suggesting prioritization of data-efficient techniques and adaptive models.

Future advancements might explore the integration of various sensor systems, amalgamating data from wearables, ambient environments, and video feeds, thereby enhancing anomaly detection accuracy. Additionally, unsupervised and reinforcement learning paradigms could introduce novel layers of sophistication, moving towards comprehensive, autonomous fall detection systems.

In summary, the paper presents a significant re-evaluation of fall detection strategies, urging the research community to embrace data availability as a primary lens. This perspective aligns fall detection methodologies more closely with real-world constraints, setting a foundation for future innovations in artificial intelligence and human activity recognition.