Silhouette Orientation Volumes for Efficient Fall Detection in Depth Videos

Area of Science:

• Computer Vision
• Human Action Recognition
• Biomedical Engineering

Background:

• Human fall detection is crucial for elder care and safety monitoring.
• Existing methods often struggle with accuracy and robustness in real-world scenarios.

Purpose of the Study:

• To develop a novel and robust method for detecting human falls using depth videos.
• To improve the accuracy of fall detection and action recognition compared to state-of-the-art techniques.

Main Methods:

• A fast and robust shape sequence descriptor, Silhouette Orientation Volume (SOV), was developed.
• The SOV descriptor was combined with Bag-of-Words and Naïve Bayes classifier for action representation and fall classification.
• Experiments were conducted on the public SDU-Fall and Weizmann action datasets.

Main Results:

• The SOV method achieved up to 91.89% fall detection accuracy on the SDU-Fall dataset using a single-view depth camera.
• This represents a 5% improvement in classification rate over existing literature.
• An overall accuracy of 89.63% was obtained for six-class action recognition, surpassing the state-of-the-art by 25%.

Conclusions:

• The Silhouette Orientation Volume (SOV) descriptor offers a highly accurate and robust approach for human fall detection and action recognition.
• This novel method demonstrates significant performance improvements over current state-of-the-art techniques.
• The SOV method shows potential for real-world applications in health monitoring and safety systems.