Convolutional Neural Network-Based Embarrassing Situation Detection under Camera for Social Robot in Smart Homes

Area of Science:

• Computer Science
• Artificial Intelligence
• Robotics

Background:

• Smart home technology deployment is hindered by privacy concerns from monitoring equipment.
• Ensuring user privacy is crucial for the acceptance of smart home systems, especially for vulnerable populations.

Purpose of the Study:

• To develop a privacy-preserving system for smart homes capable of detecting potentially embarrassing situations.
• To enhance the usability and adoption of smart home technologies for elderly and disabled care.

Main Methods:

• An improved You Only Look Once (YOLO) neural network structure was designed for feature extraction.
• A bounding-box merging algorithm based on region proposal networks (B-RPN) was developed to reduce redundancy.
• A real-time object detection algorithm based on the novel F-YOLO (Feature-YOLO) was proposed and implemented on a social robot (MAT).
• Training and validation datasets were created with 2580 and 360 images, respectively, alongside experimental test datasets.

Main Results:

• The proposed RODA-FY (Real-time Object Detection Algorithm based on F-YOLO) system achieved a recognition accuracy of 94.48% in detecting designed situations.
• Analysis explored the impact of training iterations and learning rates on prediction accuracy.
• RODA-FY demonstrated superior performance compared to Inception V3 and YOLO models in recognition accuracy.

Conclusions:

• The developed privacy detection system effectively identifies sensitive situations in smart homes, addressing key privacy concerns.
• The F-YOLO algorithm and RODA-FY show significant promise for enhancing the safety and privacy in smart home environments.
• This research paves the way for more widespread and acceptable deployment of smart home technologies in care settings.