Alex Liang

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2024-115

May 16, 2024

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2024/EECS-2024-115.pdf

In response to the growing demographic of older individuals living alone and the heightened risks of falls they face, real-time fall detection systems using surveillance videos have emerged as crucial tools for ensuring prompt assistance. This report introduces a novel real-time fall detection method that integrates learnable edges into Spatial Temporal Graph Convolutional Networks (STGCN) for enhanced accuracy in classifying human actions. Leveraging short sub-sequences of skeleton data as inputs, the proposed model achieves rapid training and inference while demonstrating robust generalization across diverse environmental conditions. The proposed method underscores its efficacy in real-world fall detection tasks. Evaluation through a devised scheme, simulating real-time video streams, validates the model's effectiveness, quantified through metrics such as accuracy, specificity, and sensitivity.

Advisors: Brian A. Barsky

---

BibTeX citation:

@mastersthesis{Liang:EECS-2024-115,
    Author= {Liang, Alex},
    Title= {Skeleton-based Fall Detection using Spatial Temporal Graph Convolutional Networks with Learnable Edges},
    School= {EECS Department, University of California, Berkeley},
    Year= {2024},
    Month= {May},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2024/EECS-2024-115.html},
    Number= {UCB/EECS-2024-115},
    Abstract= {In response to the growing demographic of older individuals living alone and the heightened risks of falls they face, real-time fall detection systems using surveillance videos have emerged as crucial tools for ensuring prompt assistance. This report introduces a novel real-time fall detection method that integrates learnable edges into Spatial Temporal Graph Convolutional Networks (STGCN) for enhanced accuracy in classifying human actions. Leveraging short sub-sequences of skeleton data as inputs, the proposed model achieves rapid training and inference while demonstrating robust generalization across diverse environmental conditions. The proposed method underscores its efficacy in real-world fall detection tasks. Evaluation through a devised scheme, simulating real-time video streams, validates the model's effectiveness, quantified through metrics such as accuracy, specificity, and sensitivity.},
}

EndNote citation:

%0 Thesis
%A Liang, Alex 
%T Skeleton-based Fall Detection using Spatial Temporal Graph Convolutional Networks with Learnable Edges
%I EECS Department, University of California, Berkeley
%D 2024
%8 May 16
%@ UCB/EECS-2024-115
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2024/EECS-2024-115.html
%F Liang:EECS-2024-115