Related Experiment Video
Updated: Oct 17, 2025
08:05
Design and Analysis for Fall Detection System Simplification
Published on: April 6, 2020
10.9K
Optimal Training Configurations of a CNN-LSTM-Based Tracker for a Fall Frame Detection System
Nur Ayuni Mohamed1, Mohd Asyraf Zulkifley1, Ahmad Asrul Ibrahim1
1Department of Electrical, Electronic and Systems Engineering, Faculty of Engineering and Built Environment, Universiti Kebangsaan Malaysia, Selangor 43600, Malaysia.
Sensors (Basel, Switzerland)
|October 13, 2021
Summary
A novel SmartConvFall system accurately detects the exact fall frame in videos using object tracking and deep learning. This automated fall detection system minimizes severe consequences by enabling timely medical intervention.
Area of Science:
- Computer Science
- Artificial Intelligence
- Biomedical Engineering
Background:
- Fall event detection is critical for minimizing severe consequences of falls, necessitating early and accurate detection.
- Existing fall detection methods often struggle with sporadic fall occurrences and false detections due to varied fall conditions.
- Precisely identifying the exact fall frame is crucial for timely response and effective treatment.
Purpose of the Study:
- To propose an automated fall frame detection system, SmartConvFall, for accurately identifying the precise moment of a fall in video sequences.
- To optimize training configurations for both the object tracking and fall frame detection modules to achieve the system's primary goal.
- To introduce a novel approach that emphasizes object tracking over mere detection for improved fall event analysis.
Main Methods:
- Developed SmartConvFall, a two-part system comprising an object tracking module using a fully convolutional neural network (CNN) and an instantaneous fall frame detection module utilizing a Long Short-Term Memory (LSTM) network.
- Evaluated various training configurations, including optimizers, learning rates, batch sizes, and region of interest, for the CNN tracker.
- Optimized LSTM model configurations by exploring feature selection and stacked layer variations to model object trajectories and detect fall frames based on vertical axis movement.
Main Results:
- The SmartConvFall system demonstrated superior performance compared to state-of-the-art trackers (TCNN, MDNET-N) with high metrics for average overlap (0.1619), robustness (0.6323), and reliability (0.7958).
- The system achieved a significantly low number of tracking failures (43 occasions).
- A three-stack LSTM model provided the lowest mean error, with an approximate one-second delay in pinpointing the exact fall frame.
Conclusions:
- The proposed SmartConvFall system offers a novel and effective approach to fall event detection by integrating object tracking with deep learning.
- The system's high accuracy, robustness, and reliability indicate its suitability for real-time applications.
- SmartConvFall has the potential to significantly improve patient outcomes by enabling prompt medical intervention, thereby reducing fall-related injuries and fatalities.
