Non-line-of-sight multi-person pose sensing

Area of Science:

• Computer Vision
• Robotics
• Human-Computer Interaction

Background:

• Human pose sensing is vital for anti-terrorism and emergency rescue.
• Existing methods primarily address line-of-sight (LOS) conditions.
• Current non-line-of-sight (NLOS) research is limited to single-person detection, hindering multi-person applications.

Purpose of the Study:

• To propose the first adaptive method for multi-person pose sensing in NLOS environments.
• To address the limitations of current research in complex, occluded scenarios.
• To enable more effective decision-making and response in critical situations.

Main Methods:

• Reconstruction of coarse 3D features from transient signals.
• Refinement of features using a neural network for improved accuracy.
• Prediction of 3D body center heatmaps and mesh parameter maps.
• Adaptive multi-person 3D mesh regression guided by body center information.

Main Results:

• Effective sensing of human poses in multi-person NLOS scenarios.
• Demonstrated performance on both simulated and real-world data.
• Successful acquisition of real-world data using a self-built NLOS system.

Conclusions:

• The proposed method enables effective multi-person pose sensing in NLOS environments.
• The developed simulation pipeline and real-world dataset are valuable resources for the research community.
• Public release of the dataset will foster further advancements in NLOS human pose sensing.