EgoFall: A First-Person View Fall Detection System.

Area of Science:

• Gerontology
• Computer Science
• Biomedical Engineering

Background:

• Falls are a significant health risk for older adults, with millions experiencing them annually.
• Existing fall detection systems often depend on fixed surveillance cameras, which have limitations in placement and line of sight.
• There is a need for unobtrusive and effective fall detection solutions for elderly care.

Purpose of the Study:

• To develop and evaluate EgoFall, a novel wearable camera-based system for fall detection using first-person imagery.
• To assess the feasibility of implementing EgoFall on resource-constrained hardware like a Raspberry Pi.
• To compare the performance of different machine learning models for fall event identification.

Main Methods:

• EgoFall employs Oriented FAST and Rotated BRIEF (ORB) for keypoint extraction from first-person images.
• Optical flow methods are utilized to compute speed and direction features indicative of falls.
• A Support Vector Machine (SVM) model was trained and evaluated for fall event classification.

Main Results:

• The EgoFall system was successfully implemented on a Raspberry Pi 3B.
• Among the evaluated models (SVM, KNN, DT), SVM demonstrated the highest performance.
• The SVM model achieved an accuracy of 84.9% in detecting fall events using the RUG-EGO-FALL dataset.

Conclusions:

• EgoFall presents a viable alternative to traditional surveillance-based fall detection systems.
• Wearable, first-person camera systems can effectively detect falls, enhancing elderly care safety.
• Further research can optimize machine learning models and hardware for improved fall detection accuracy and efficiency.