Deep Learning for Fall Detection: Three-Dimensional CNN Combined With LSTM on Video Kinematic Data

Area of Science:

• Computer Science
• Biomedical Engineering
• Artificial Intelligence

Background:

• Fall detection is crucial for public healthcare, enabling timely medical aid.
• Current video-based methods often require extensive datasets and are sensitive to image quality.
• Collecting real fall data is challenging, leading to reliance on limited simulated datasets.

Purpose of the Study:

• To develop an effective fall detection system using 3D CNN and LSTM with spatial attention.
• To overcome the limitations of large dataset requirements and image quality dependency in existing methods.
• To improve the accuracy and robustness of automated fall detection.

Main Methods:

• A 3D convolutional neural network (3-D CNN) was employed to extract motion features from video kinematic data.
• A long short-term memory (LSTM) network with a spatial visual attention mechanism was integrated for region localization.
• The 3-D CNN was initially trained on the Sports-1M dataset, then combined with LSTM for fall detection model training.

Main Results:

• The proposed scheme achieved 100% accuracy on a fall detection benchmark.
• The method demonstrated superior performance compared to existing approaches on various activity databases.
• The system effectively extracts temporal motion features and spatial information for accurate fall identification.

Conclusions:

• The developed 3-D CNN and LSTM-based method offers a robust and accurate solution for fall detection.
• This approach circumvents the need for large fall-specific datasets, making it more practical.
• The system shows significant potential for real-world healthcare applications, enhancing patient safety.