密度クラスタリング誘導型ガウスモデルフィッティングに基づく単一光子減霧撮像法

In strong smoke-scattering environments, single-photon counting lidar faces the dual challenges of intense scattering noise and low echo photon counts, which result in extremely low signal-to-noise ratios (SNRs) and consequently severely limited smoke-penetrating imaging performance. The residual noise photons persist after the Gamma model fitting of the backscattering peak in smoke environments and interfere with the accuracy of the target Gaussian model fitting. To address this issue, this work proposed a smoke-penetrating 3D imaging algorithm based on DBSCAN residual clustering-guided Gaussian model fitting. First, a Gamma model was constructed to fit the echo photon data, provide an initial estimate of the target echo position, and separate the backscattering peak. Then, a DBSCAN density clustering algorithm was designed to analyze the fitting residuals and effectively identify and remove residual noise photons to precisely isolate the target signal peak. Finally, Gaussian model fitting was performed on the filtered target photons to achieve high-precision depth estimation. The proposed method effectively separates the signals from noise in smoke-scattering environments using DBSCAN density clustering. Experimental results on the Middlebury simulation dataset show that the proposed algorithm achieved lower root mean square error (RMSE) and higher structural similarity (SSIM) in both the Art and Bowling scenes across varying smoke particle sizes (0.4,1,10 and 20 µm) compared to conventional algorithms. Furthermore, real-world outdoor experiments under rainy and foggy conditions further confirmed its reconstruction capability. These improvements significantly enhanced the quality and accuracy of imaging in smoke-scattering environments.