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

Conventional object tracking is restricted to targets within the direct field of view. Recent advances in non-line-of-sight (NLOS) imaging provide promising solutions for dynamic target tracking in occluded environments. Although most existing methods focus on corner-based three-bounce scenarios, two-bounce NLOS configurations, such as those found in corridors and tunnels, have not been fully explored. In this work, we propose a time-multiplexed detection system that captures dual-view shadow images under alternating laser illumination. The captured image sequences are processed by the dual-view collaborative perception network (DCPNet), which calculates spatio-temporal trajectories by integrating cross-view spatial consistency and single-view temporal coherence within a multi-stage hierarchical architecture. To overcome the lack of real-world data, we construct a large-scale synthetic dataset using physics-based rendering and further validate the model's generalization through real-world experiments. Extensive evaluations demonstrate that our method achieves an RMSE of as low as 0.026 on the simulated dataset. After brief fine-tuning on real-world data, our method can achieve high-accuracy real-time tracking under varying motion patterns and shows good generalization across real-world scenarios.