Semantic segmentation of 3D LiDAR data using deep learning: a review of projection-based methods

LiDAR sensor is an active remote sensing sensor that is increasingly used to capture 3D information of real-world objects. Real-time decision-making applications such as autonomous driving heavily rely on 3D information to navigate an urban environment. LiDAR data processing is, however, very complex and resource-intensive. Deep learning on point cloud is a recent advancement that is aimed to extract 3D information. Deep learning implementations include procedures where raw points are fed to neural networks and converted to 3D voxels. Individual voxels are fed to 3D convolutional layers and techniques that transform the 3D points into 2D images and utilize the well-established 2D CNNs. Of these, the two former methods are majorly reviewed, while the projection-based methods are less reviewed although the technique is widely used in numerous applications. To fill the gap, this paper examines the existing literature on projection-based methods by detailing the recent progress made. Identifying the state-of-the-art methodology and summarizing the important interventions are among the significant tasks covered in this paper.