Fast three-dimensional scanning imaging using single pixel single photon lidar under extremely low signal to noise ratio conditions

Single photon lidar achieves high temporal resolution through the integration of time correlated single photon counting (TCSPC) and time of flight (ToF) techniques, enabling highly sensitive photon detection. It has become a key technology in active target detection, autonomous driving, and 3D imaging. However, under strong background noise conditions, its performance is severely limited by low signal to noise ratio (SNR) and weak echo signals. To address these challenges, we present an enhanced PRS-NET deep learning architecture based on hardware software codesign. Simulation results demonstrate that even at an SNR as low as 0.02, the reconstruction root mean square error (RMSE) reaches only 0.0153 m, with a reconstruction accuracy of 97.98% under the accuracy@δ metric with δ = 1.03, i.e., 97.98% of the depth estimates have a relative error smaller than 3%. Using this algorithm integrated with a single photon lidar system, we conducted a rapid 3D scanning experiment during daytime over a 7.7 km urban atmospheric path. Under an average signal photon count per pixel below one (i.e., the mean number of detected signal photons per pixel across the 1024 × 1408 image is < 1) , we achieved high resolution imaging at 1000 × 1410 pixels, with a single point accumulation time of just 0.2 ms and a ranging frame rate of 5 kHz@7.7 km exceeding current state of the art performance. The proposed method significantly improves detection efficiency under extreme noise conditions, thereby advancing the practical deployment of single pixel scanning imaging and broadening its applicability in future real-world scenarios.