Abstract:Geiger-Mode Avalanche Photodiode (Gm-APD) imaging is typically accompanied by significant noise interference, which poses challenges to the accurate extraction of target information. To address the issues of low signal-to-noise ratio (SNR) and extremely weak echo signals, this paper proposes a spatiotemporal joint denoising method based on 3D convolution. The algorithm leverages the high correlation of Gm-APD LiDAR echo signals in both the temporal and spatial domains, utilizing a designed 3D convolution kernel to fully extract and integrate temporal and spatial information. The method first counts the histogram of the trigger frequency of each pixel across multiple consecutive frames in the temporal domain. Then, it employs the designed 3D convolution kernel to extract temporal and spatial information, obtaining an optimized trigger histogram, and distinguishing target points from noise points through threshold segmentation. Experiments are conducted on actual Gm-APD LiDAR datasets. The experimental results demonstrate that, compared to the peak-threshold method, the proposed method achieves a 39.96% improvement in target recovery and a 9.15-fold increase in SNR under the condition of 200-frame fusion denoising, indicating enhanced denoising performance.