Abstract:Gm-APD Lidar can realize single-photon level detection and fast long-distance three-dimensional imaging, which has important application value in many fields. However, there is currently no large-scale public dataset constructed by the data of this Lidar, which limits the application of deep learning-based algorithms on this type of data. Therefore, proposed a two-stage data generation model. First, on the basis of the DCGAN network, a convolutional attention module and a skip connection structure were designed for the generator, and the output of discriminator was modified to the PatchGAN architecture and the parameter spectra of the two were normalized to obtain a high-quality intensity image; after that, use the distance image and binarized intensity image to train the distance image generation model with pix2pix as the original network, learn the dependency between pixels and reduce the checkerboard effect by introducing the self-attention and pixel shuffle module, and finally get the intensity image constrained distance image. The performance of the data generation model proposed in this paper on the self-built dataset have obtained the best scores under various evaluation indicators, and the quality of generated data is far better than the existing model. The dataset expanded by this model effectively improves the detection accuracy of the 3D object detection network PointRCNN. The method proposed in this study proposes a feasible solution for Gm-APD LiDAR data generation.

Data Generation Model Based on Gm-APD LiDAR Data