Real-time two-bounce non-line-of-sight object tracking via dual-view collaborative perception network

Area of Science:

• Computer Vision
• Robotics
• Optical Imaging

Background:

• Conventional object tracking is limited to line-of-sight scenarios.
• Non-line-of-sight (NLOS) imaging offers solutions for tracking in occluded environments.
• Two-bounce NLOS configurations, common in corridors, remain underexplored.

Purpose of the Study:

• To develop an effective object tracking system for two-bounce non-line-of-sight (NLOS) environments.
• To address the limitations of existing NLOS tracking methods.
• To enable real-time dynamic target tracking in occluded settings.

Main Methods:

• A time-multiplexed detection system capturing dual-view shadow images under alternating laser illumination.
• Processing image sequences with the dual-view collaborative perception network (DCPNet).
• Utilizing a multi-stage hierarchical architecture integrating cross-view spatial consistency and single-view temporal coherence.

Main Results:

• A large-scale synthetic dataset was created using physics-based rendering.
• The proposed method achieved a Root Mean Square Error (RMSE) as low as 0.026 on the simulated dataset.
• Real-world experiments demonstrated high-accuracy real-time tracking and good generalization after fine-tuning.

Conclusions:

• The dual-view collaborative perception network (DCPNet) effectively enables non-line-of-sight (NLOS) object tracking.
• The method shows promise for real-time applications in occluded environments like tunnels and corridors.
• Validation through synthetic data and real-world experiments confirms the system's robustness and generalization capabilities.