Imitating Emergencies: Generating Thermal Surveillance Fall Data Using Low-Cost Human-like Dolls.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Outdoor fall detection is crucial for public safety but lacks sufficient training data.
• Collecting real-world fall data is challenging, dangerous, and ethically complex.
• Existing automated surveillance methods often overlook fall detection in diverse outdoor environments.

Purpose of the Study:

• To propose a safe and cost-effective method for generating fall data for outdoor environments.
• To simulate real-world emergency scenarios using thermal imaging of a falling object.
• To validate the utility of simulated fall data for training deep learning models.

Main Methods:

• Utilized a low-cost rubber doll falling in a harbor to simulate accidental falls.
• Captured thermal images to achieve human-like thermal signatures on the doll.
• Measured and verified the stability of thermal signature changes over time.
• Applied a state-of-the-art object detector trained on real people to the captured doll data.

Main Results:

• Simulated fall videos exhibited motion and appearance similar to real human falls.
• Achieved an average confidence score of 0.730 for pedestrian detection using doll data.
• Demonstrated comparable performance to using actual footage of people falling (0.761 confidence score).
• Confirmed the stability and reliability of thermal signatures from the falling doll.

Conclusions:

• The proposed method offers a viable and safe alternative for generating fall detection datasets.
• Thermal imaging of falling dolls can effectively substitute real human fall footage for training AI models.
• This technique enhances the development of robust outdoor fall detection systems for public safety applications.