Robust fall detection in video surveillance based on weakly supervised learning

Area of Science:

• Computer Science
• Artificial Intelligence
• Biomedical Engineering

Background:

• Fall event detection is crucial in medicine and health, with vision-based methods offering non-contact advantages.
• Existing vision-based fall detection relies heavily on supervised learning, demanding extensive data annotation.
• This creates a bottleneck in terms of time and resources for developing effective fall detection systems.

Purpose of the Study:

• To propose a novel fall detection method that overcomes the limitations of supervised learning by utilizing weakly supervised learning.
• To develop a dual-modal network capable of learning fall events from weak labels, thereby reducing annotation effort.
• To enhance detection accuracy and efficiency in vision-based fall event monitoring.

Main Methods:

• A weakly supervised learning approach using a deep multiple instance learning framework is employed.
• A dual-modal network architecture is designed to process information from two distinct streams.
• A novel dual-modal fusion strategy is introduced to integrate outputs from both network streams for final detection.

Main Results:

• The proposed method successfully learns fall events using weak labels, significantly reducing annotation requirements.
• Experimental results on benchmark and novel datasets demonstrate the method's superiority over state-of-the-art approaches.
• The dual-modal fusion strategy effectively integrates information for robust fall detection.

Conclusions:

• The weakly supervised dual-modal network offers a more efficient and effective solution for vision-based fall detection.
• This approach alleviates the burden of fine-grained data annotation, making fall detection systems more accessible.
• The method shows significant promise for real-world applications in healthcare and elderly care monitoring.