Image-free single-pixel keypoint detection for privacy preserving human pose estimation.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Computer vision is widely used but raises privacy concerns due to human image data.
• Existing methods often require detailed visual information, posing security risks.

Purpose of the Study:

• To propose an image-free human keypoint detection technique.
• To address data security and privacy issues in human-centric computer vision tasks.
• To develop a resource-efficient and privacy-preserving solution.

Main Methods:

• Utilized coded illuminations and a single-pixel detector for keypoint detection.
• Developed an encoder-decoder network to optimize illumination patterns and keypoint prediction.
• Trained and validated the model using Leeds Sport Dataset and COCO Dataset.
• Incorporated EfficientNet backbone to reduce inference time.

Main Results:

• Achieved 91.7% average precision in simulation.
• Demonstrated 88.4% average precision experimentally at a sampling rate of 0.015.
• Reduced inference time from 4 seconds to 0.10 seconds.
• The method operates on a 1D sequence without image reconstruction, ensuring privacy.

Conclusions:

• The proposed image-free method offers significant privacy protection and resource efficiency.
• It successfully detects human keypoints at an ultralow sampling rate.
• Potential applications include clinical monitoring, construction site surveillance, and home service robots.