Three-Dimensional Posture Estimation of Vehicle Occupants Using Depth and Infrared Images.

Area of Science:

• Computer Vision
• Robotics
• Human-Computer Interaction

Background:

• Pose estimation is vital for autonomous vehicle safety and user experience.
• Traditional color-image methods struggle with 3D data, occlusion, and lighting.
• Depth images offer 3D information and lighting invariance but lack sufficient research for pose estimation.

Purpose of the Study:

• To develop a novel 3D human pose estimation method using only depth and infrared (IR) images.
• To address the challenge of limited annotated depth data for training pose estimation models.
• To enhance passenger safety in semi- and fully autonomous vehicles through accurate occupant posture detection.

Main Methods:

• A three-stage fine-tuning process using simulated, approximated, and limited manually annotated data.
• Leveraging both depth and IR image data for robust 3D pose estimation.
• Training a model for vehicle occupant posture detection.

Main Results:

• Achieved accurate 3D pose estimation with a median error under 10 cm across all joints.
• Required fewer than 100 manually annotated samples for effective model training.
• Demonstrated the first work focusing on vehicle occupant posture detection using only depth and IR data.

Conclusions:

• The proposed method is feasible and effective for 3D human pose estimation using depth and IR images.
• This approach significantly reduces the need for extensive manual annotation.
• Paves the way for improved passenger safety systems in autonomous vehicles.