An Approach to Translate Synthetic Lidar Point Cloud Data Similar to Real Lidar Data and Evaluate Using Deep Learning Techniques

As autonomous driving technology advances, the need for 3D LiDAR point cloud data has grown, particularly for handling unpredictable daily travel scenarios. However, it typically requires large annotated datasets for training, posing challenges in the 3D semantic segmentation models which is time-consuming and expensive. To address this, synthetic data from simulators is considered, but it lacks realism compared to actual LiDAR data. This study introduces a method that translates synthetic point clouds to closely resemble real data, enabling their use in real-time applications with limited actual LiDAR data. The proposed method combines Point Cloud Translation (PCT) and Domain Adaptation (DA) techniques, along with the state-of-the-art Cylinder3D segmentation network for 19 classes. Along with the above generation, comparative evaluations of various mixing techniques on synthetic and domain-translated data determine the optimal ratios for 3D semantic segmentation when combined with real data. SynLiDAR serves as the synthetic source data, and Semantic Kitti serves as the target data. By combining synthetic and domain-translated data with real data, the proposed model achieves mean Intersection over Union (mIOU) scores of 63.954 and 64.578, surpassing the mIOU of real data alone using the Cylinder3D segmentation network.