Walking stability prediction for pedestrians using gait energy images and hybrid deep and few-shot learning models

Clinical Snapshot

Key Findings

• P Value: Not reported

• Effect Size: YOLOv8-cls achieved 96.92% overall accuracy (97.14% for Loss-of-Balance anomalies; 94.67% for Active Motion anomalies) versus Linear SVM baseline of 75.38% — an absolute improvement of 21.54 percentage points. LODO cross-validation mean accuracy: 83.28%

• Primary Outcome: Binary classification of stable versus unstable pedestrian walking patterns using Gait Energy Images across twelve machine learning and deep learning methodologies

• Nnt Or Sensitivity: Sensitivity and specificity not reported. No NNT calculable. MobileNet edge deployment latency: 4.2 ms. SAM + MediaPipe preprocessing yielded 12.50 percentage-point absolute accuracy improvement over baseline preprocessing pipeline

• Confidence Interval: Not reported for any metric

Clinical Application

MobileNet's 4.2 ms inference latency supports real-time edge deployment in principle. However, clinical implementation requires camera infrastructure, integration with electronic medical records or nurse call systems, privacy governance frameworks, and clinical workflow integration — none of which are addressed. The SAM + MediaPipe preprocessing pipeline adds computational overhead not fully quantified for end-to-end system latency. Australia's aged care sector — encompassing approximately 900 residential aged care facilities and a growing home care sector — represents a plausible deployment context. The Royal Australian College of General Practitioners (RACGP) and the Australian Commission on Safety and Quality in Health Care (ACSQHC) have established fall prevention frameworks (e.g., Preventing Falls and Harm from Falls in Older People guidelines) that emphasise validated, multifactorial risk assessment. This technology is not TGA-regulated as a medical device in its current research form but would require TGA Software as a Medical Device (SaMD) assessment prior to clinical deployment. PBS listing is not applicable. No Australian population data are included in the study. Privacy considerations under the Australian Privacy Act 1988 and aged care-specific surveillance regulations would require careful navigation before deployment in residential or community settings. The study targets elderly individuals at risk of falls and individuals in smart surveillance environments. However, the dataset does not include clinically characterised elderly fallers, patients with Parkinson's disease, stroke survivors, or other high-risk groups. Application to any specific clinical population requires prospective validation in that population.

Abstract

The prediction and recognition of unstable human walking patterns are of high importance for active video surveillance, smart environments, and assistive healthcare, particularly for fall detection in the elderly. This research investigates the utility of Gait Energy Images (GEIs) combined with deep learning, vision transformers, and few-shot learning architectures to enhance the classification of stable and unstable pedestrian walking patterns. We evaluate and compare twelve methodologies: six classical or feature-based machine learning models (Linear SVM, HOG + SVM, LBP + RBF-SVM, Random Forest, XGBoost, and an adapted GaitSet baseline), three deep learning models (MobileNet, Vision Transformer (ViT), and YOLOv8-cls), and three episodic few-shot learning techniques (Prototypical, Matching, and Relation Networks) under data-scarcity regimes. To facilitate this evaluation, we introduce the Unstable and Stable Walking Pedestrian (USWP) dataset, constructed by fusing and harmonizing sequences from seven public action recognition databases, containing 3250 unique GEIs with a subject-independent evaluation protocol to prevent identity-based domain leakage. Our experiments demonstrate that the YOLOv8-cls model achieves an overall accuracy of 96.92% (97.14% on Loss-of-Balance anomalies and 94.67% on Active Motion anomalies), significantly outperforming the conventional Linear SVM baseline (75.38%) and MobileNet (91.08%). Conversely, Relation Networks exhibit lower few-shot performance (71.08%) due to optimization complexities in learning similarity metrics from sparse data. Leave-One-Dataset-Out (LODO) cross-validation reveals an average generalization accuracy of 83.28%, indicating significant domain bias across source databases and underscoring that within-dataset evaluations overestimate real-world generalization. Computational complexity analysis shows that MobileNet provides an optimal trade-off for real-time edge deployment (4.2 ms latency), while preprocessing ablation studies demonstrate that integrating the Segment Anything Model (SAM) with MediaPipe-derived Regions of Interest (ROIs) yields a 12.50% absolute improvement in accuracy by eliminating background noise.

References

1. 1.Taha, M., Fares, A., Yamaguchi, H., & Zaky, A. B. (2026). Walking stability prediction for pedestrians using gait energy images and hybrid deep and few-shot learning models. Scientific Reports. PubMed ID: 42469337.